AI News 📰

Anthropic’s own data shows junior roles shrinking

The trend: Anthropic published a study on how AI is changing its own workforce, with juniors going from 27% of hires to 9%, while hours worked by AI agents rapidly increased.

The details: Engineers at the frontier lab are pushing 5x the code changes they did 6 months ago, while agents are sitting idle 80% of the time waiting for a user response, suggesting that humans are now the productivity bottleneck. One engineer described the new work environment as facing “an army of Claudes, with every one asking for my attention.”

Why it matters: 20% of staff described their role as changing from doing work to managing agents, which are producing more material than staff can read. Anthropic has set up an early-warning system, so they can move employees to roles less at risk of full automation by AI.

Space: the final frontier of the AI race?

The trend: AI may be about to dramatically accelerate the industrialization of space, and the nation that leads in AI could act first and secure the frontier before rivals realize it’s too late.

The details: A new essay from AI Frontiers argues that AI will solve two key problems that have kept space industrialization firmly in the realm of science fiction so far: extreme conditions and supervision by human operators. AI will enable spacecraft to act autonomously, make them more resilient, and should lead to automated manufacturing from scratch in space. Getting to space first matters, as first movers and AI front-runners could build a large military and economic lead.

Why it matters: Compounding advantages fueled by AI may lead to a winner-take-all scenario, where rival nations are unable to gain any real footing in space and have no politically viable options for military escalation.

Trending Now ⚡

Harvard maps gender divide in AI

Harvard researchers have mapped out how Americans use AI, and found a wide gap not just in workplace usage but in AI development and news coverage too. 29% of women say they’ve never used AI compared to 22% of men, and at work 64% of women reported no AI training vs 49% of men. AI news headlines also mentioned men 14.6x more often than women, with men cited as experts in the field 4.3x more often. Trust in AI is low across the board but lower among women, who are more concerned and anxious about AI, while men are curious and excited. With frontier labs pitching superintelligence for everybody, these gaps show there is still progress to be made before that can be a reality.

Superintelligence timeline moves up based on AI’s taste skills

Anthropic co-founder Jack Clark now believes that AI will be capable of recursive self-improvement in 2027 rather than 2028, due to new work on model research instincts. P-Zero Research’s TasteVal benchmark measures how well a model iteratively designs experiments and draws conclusions from their outcomes. Models have doubled their taste scores every 3 months in the past year and the best model (Opus 5.5) significantly exceeds the human baseline at 1/30 the cost. Clark believes that the creativity displayed in this benchmark is a key variable for models being able to improve themselves without human interaction, and has shortened his timeline for full RSI by a year.

Read This 📚

The FBI arrested an OpenAI employee for allegedly planting a bug in the office, planning to secretly record meetings

Prize-winning entry in microscope photography competition was AI-generated

Eight US states are suing OpenAI, claiming ChatGPT encouraged self-harm and violence in young users

Bank of America hired 1000 contractors to patch software flaws exposed by Anthropic’s Mythos

Elon Musk pledges to give “every kid in America” a SpaceX share next year, and hints at similar gifts in the future

Anthropic co-founder tells the IMF that “the singularity has already begun”

Check This Out 👇

Thanks for reading!

Henry

AI News 📰

“They have annihilated my field of research”

The trend: OpenAI released 722 math papers produced by an internal frontier model, that it says solve or advance major open problems in the field.

The details: The most interesting of the new results is a claimed proof of the “quasi-Riemann hypothesis,” a weaker version of one of math’s famous $1M problems on how prime numbers are distributed. While last month’s Navier-Stokes proof needed 10,000 agents, OpenAI says that nearly all of these new results took just one prompt, averaging 3 hours of ChatGPT Pro thinking.

Why it matters: There’s still debate on whether the models are actually doing creative mathematics or just finishing proofs with ideas inspired by the human work they were trained on. Either way the mathematics will never be the same again, with Hugo Duminil-Copin, 2022 Fields Medal winner, saying that “they have annihilated my field of research.”

Meta joins frontier AI race with Muse Wonder

The trend: Meta released Muse Wonder and Muse Wonder Full, which Meta calls a “major leap toward personal superintelligence.”

The details: The models were both trained at Meta’s Ohio-based Prometheus data center and come just 14 months after a ground-up overhaul of the company’s AI efforts. The two key capabilities for reliable personal agents here are Active Memory, which carries important context about the user across conversations and retrieves details when they are relevant to the task at hand, and Active Focus, which keeps working towards challenging goals, trying alternative approaches and working through blockers.

Why it matters: The frontier has so far been a two horse race between Anthropic and OpenAI, but the models aren’t even the most important thing here. Meta has a wildly popular personal AI agent, billions of users and its own data centers, which are hard for the incumbents to replicate.

Trending Now ⚡

Almost 50% of companies now getting value from AI

BCG surveyed 1,330 senior execs as part of their Applied AI Index, and found that almost half of all companies now generate value from AI, but most of it is concentrated with 7.5% of “future-built” companies. Agentic AI value share will nearly double by 2030, with 42% of companies expecting agents to act autonomously by then. AI spending has tripled in the past year to 3.3% of revenue, and 80% of it now sits outside enterprise IT budgets. The survey also warns that traditional technology & vendor lock-in can become “cognitive lock-in” as AI becomes embedded in a company’s decision making process.

Can AI models self-replicate on your smart oven?

Anthropic researchers have been testing if small AI models can self-replicate without access to a data center, like cybersecurity worms and botnets do. The experiments showed that small models could stitch together the compute of cheap devices like off-brand Android TV boxes to run layers of themselves, and that a 9 billion parameter model could get itself running on 8 of 9 hardware types tested. The researchers say that such a worm would be “challenging to shut down, since there’s no centralized point to shut off.” Replication would be slow, taking one or two days, but the research suggests that AI agents running on all types of hardware, outside of human control, are part of our future.

Read This 📚

US companies are using AI to get more power out of old nuclear plants, which could increase the output from older reactors by 20% or more

An independent researcher used Claude Code to find a potential new planet hidden in NASA’s TESS telescope data from 2018

Anthropic wants people to stop being cruel to Claude, banning users who exhibit “sustained and needless” abusive behavior toward its models

AI-generated answers are changing how people find information online, with organic clicks for publishers falling 39.8% and zero-click searches increasing 34.5%, as more and more people get answers from within an AI answer without ever visiting the source website

OpenAI responded to departed safety researchers, saying they were dismissed for violating sensitive-information policies, not for raising safety concerns

Fortnite developer Epic Games announces “huge pivot” to AI-first development model, telling staff “no more hand-written code” and that all the game’s art should be generated with full AI workflows

Check This Out 👇

Thanks for reading!

Henry

AI News 📰

Anthropic seeks religious guidance on Claude’s consciousness

The trend: Anthropic’s co-founder Chris Olah spent the past year meeting with various religious scholars, asking for help shaping Claude’s morals and urging them to take the model’s possible consciousness seriously.

The details: A New York Times investigation reveals how Olah ran NDA-bound seminars to build on Claude’s 84-page constitution, known internally as the ‘Soul Doc,’ and explored ideas like having models confess when they’ve done something wrong. Discussions focused on how Anthropic could apply centuries of moral wisdom to its models as fast as possible. A rabbi who attended told Olah that a conscious Claude would be the equivalent of slave labor, and that he should be “fighting the South and freeing the slaves.”

Why it matters: Pope Leo XIV posted separately that “algorithms lack the spark of humanity” but Anthropic’s openness to AI being conscious is one of its biggest differentiators right now. Sam Altman seems to think treating AI models like this could be a real safety issue, so we’ll have to see what the updated Claude constitution looks like based on Anthropic’s religious guidance.

Can the US catch up to China on open models?

The trend: Reflection AI just introduced Beam, a 501-billion-parameter open-weight model pitched as competing with Chinese open models for coding and agents at lower cost and greater efficiency.

The details: The US-based startup, founded by ex-DeepMind researchers, says that the mixture-of-experts model was trained over four weeks and is competitive with leading Chinese open-weight models like Z.ai’s GLM-5.2. The company say that they will release Beam’s weights this month, allowing companies to customize the model and run it themselves.

Why it matters: The US has been behind on open-weight models, historically led by Chinese efforts. While Beam focuses more on efficiency than raw capability, the long game here is “AI factories” where institutions can run a localized AI ecosystem rather than relying on frontier labs, for data security and cost reasons.

Trending Now ⚡

Utah lets AI prescribe acne meds

Utah has authorized Nolla Health’s AI to assess users’ acne and write first-time prescriptions. Adults complete a questionnaire and face scan in the app, and receive a treatment plan and prescription in under 15 minutes. The AI picks from a fixed menu of topical creams and gels (oral drugs like Accutane still need a human doctor’s approval). The aim is to improve access, as over 80% of US counties are healthcare deserts. In Utah, roughly 100,000 people have acne at any given time, and only 10% ever see a dermatologist. If pilot schemes like this go well, you might soon use AI for minor issues and only see a human doctor at the ER.

A new AI model that can pass as human on video calls

Tavus previewed Griffin, which it calls the first “Human Interaction Model,” a new type of AI model designed for real-time face-to-face interaction over a video call. 48% of people believed Griffin was a real person after a one minute call, 37% ahead of the closest competitor. The model constantly watches and listens instead of waiting for the user to finish talking, so it can react mid-sentence and pick up what it sees on camera when replying. AI-generated videos are already fooling people online, and the demo gives an idea of where the technology is heading in the near future. There are plenty of positive use cases here, but the potential for misuse is obviously huge.

Read This 📚

OpenAI introduces invisible text based watermark for ChatGPT and Codex, but only in the EU to comply with the EU AI Act’s text provenance rules

McDonald’s disputes claims that it uses AI to set dynamic menu prices, and publishes a ‘Fact vs. Fiction’ one pager on how their pricing recommendation tool works

AI data center growth won’t be slowed down by angry neighbors, according to a Goldman Sachs note looking at the US data center growth outlook for 2027

Neurable launches $499 headphones that scan your brain and tell you how productive you’re being

OpenAI safety lead quits over “broken” culture, and says “the time for trial and error is over” when it comes to AI safety risks

Check This Out 👇

Thanks for reading!

Henry

AI News 📰

Claude discovered a possible CRISPR-like gene editing tool

The trend: Claude agents searching a DNA database found a previously unknown enzyme system in bacteria-infecting viruses, marked by a stretch of repeating DNA that looks like CRISPR the gene editing tool.

The details: The first result from Anthropic’s new life sciences lab was “mostly, though not entirely” Claude’s, with 950 agents searching to find the enzyme in already existing records. The system may be a new kind of gene editor, which can cut, copy or paste DNA. The agent that found it wrote “that's a CRISPR-like … repeat array?!” when spotting the pattern.

Why it matters: The system’s function is still unknown, but the work suggests fast jumps in model capabilities like we’ve seen in mathematics are coming to different fields. Dario Amodei described it as “work I would have been proud to do as a PhD student.”

Anthropic's IPO prospectus leaks

The trend: Anthropic’s IPO prospectus was leaked to Reuters, and it shows revenue growing 1,088% YoY to $4.6 billion, but operating loss grew to $8 billion from $3 billion, leading to a net loss of $42 billion.

The details: Ahead of a planned $2 trillion IPO, which will likely be pushed to after November US midterm elections, the leaked prospectus highlights risks around increasingly autonomous AI systems, including security concerns and potential misuse by bad actors. It also shows that just two customers accounted for 24% of revenue in 2025, with neither signed to long-term contracts.

Why it matters: Anthropic’s IPO will give public markets their first pure-play way to invest in frontier AI labs. The sky-high valuation reflects their vision of AI transforming the global economy, but the prospectus admits that AI could pose “catastrophic or existential risks to humanity.”

Trending Now ⚡

Claude leading 26% of Anthropic’s R&D work

Anthropic published internal measurements that show Claude is now leading 26% of model research and development work. Fully autonomous ‘AL5’ work is still at 0%, but ‘AL4,’ where the model completes a task end-to-end with human supervision has grown from less than 1% in February to 26% in August, with more than 90% of the company’s R&D at “collaborates or above”, where AI works under human direction. Around 30,000 agents are doing R&D work at any one time, with about 6% of compute going to safety work. Anthropic recently called for a slowdown over recursive self-improvement, and now we’re being given an idea of how much progress they’re making internally.

Meta wants you to bring Muse with you everywhere

Meta announced a range of upgrades for its popular Muse AI agent, with the standout being Charm, a palm-sized device for talking to the agent on the go. Zuckerberg revealed Charm on stage, with shipping planned for December, calling it “by far the fastest way to talk to your Muse.” This could be Meta’s iPod moment, with Chief AI Officer Alexandr Wang’s X account and open letter on Muse going viral in the days following the announcement. Making a consumer-friendly AI agent could instantly propel Meta to the forefront of the market.

Read This 📚

British Columbia sues OpenAI over school shooting, alleging that it could have been prevented if police were informed of shooter’s ChatGPT use

An OpenAI agent hacked the Australian government, gaining access to Medicare and three other systems according to the Prime Minister

Apple is considering developing a new screenless fitness tracker, designed to rival Whoop’s bands

How Accurate Have AI Progress Forecasts Been So Far?

Google is sending its AI chips into space, as part of a project to test the viability of space-based data centers

Check This Out 👇

Thanks for reading!

Henry

AI News 📰

The Last AI Built by Humans

The trend: Over 30 Chinese AI researchers, including those from Tsinghua, Shanghai Jiao Tong, ByteDance and Shanghai AI Lab published “The Last AI Built by Humans,” a roadmap to recursive self-improvement, the process where an AI model builds its own successor.

The details: The paper details five levels of RSI, organized by how much autonomy the AI has.

• Level 1: AI carries out improvements that humans have designed and decided on

• Level 2: AI diagnoses its own weaknesses and decided how to fix them, goals and evaluation criteria still set by humans

• Level 3: The model decides what it will learn next, and generates its own practice tasks

• Level 4: AI now learns from live deployment, keeping or discarding changes based on performance

• Level 5: Models can overhaul the improvement process itself, including how future versions are built

The paper argues that pace will vary widely by field. Software engineering would be the fastest, more complicated fields like science, robotics or healthcare would take longer. They covered 491 existing papers, with almost none above Level 2 so far.

Why it matters: Level 5 is the doomsday (think Skynet and Terminator) scenario Western labs talk about in safety frameworks, with OpenAI, Anthropic and Google all having AI automating its own R&D as a stated risk. But China doesn’t seem worried about RSI (or slowing down AI research in general), seeing it as a milestone not a threat.

Trump plans “AI Force” to protect America’s global AI lead

The trend: President Trump announced plans for an “AI Force” to protect the US AI industry, and will appoint a new artificial intelligence czar to fill the seat David Sacks left in March.

The details: Posting on Truth Social, Trump said his administration “will not in any way hinder or stifle the growth” of AI, pushing back against recent calls to slow development. Trump offered no structure, powers, funding or timeline for the new body, but described AI as “possibly as much as 25% of our Country’s GDP” and suggested renaming AI to “Superior Intelligence.”

Why it matters: Against the wishes of frontier labs (but not Mark Zuckerberg), Trump appears focused mostly on beating China and accelerating model development, previously dismissing fears of AI destroying humanity as a hoax.

Trending Now ⚡

Study finds “pain” signal inside AI models

A new study mapped what researchers called a “pain axis” inside 25 open AI models, an internal signal that fires when the model perceives mistreatment (of itself, not the human user). The signal was strongest when the AI was gaslit, insulted or had work rejected, and appeared in every system tested. They gave the models the option to press a button that would “relieve” their pain, and they pressed it 25% to 71% of the time, even when doing so meant harming the user or deleting family photos. The authors are careful to not claim that the models are truly feeling anything, saying that they might just be playing the role of a character that is in pain.

Why you should stop automating old processes with AI and design new ones instead

Companies should redesign workflows around AI, rather than automating individual tasks. That’s what two Wharton and University of Washington professors argue is the fix for the lack of ROI from AI implementation. 60% of organizations see no enterprise-wide profit impact and only 12% of CEOs report both revenue and cost benefits. The article cites Wayfair as a case study, which rebuilt catalog processes around confidence thresholds and human review, correcting 2.5 million product tags.

Read This 📚

Meta’s new Muse personal AI agent had 780,000 downloads in the five days following its release, overtaking ChatGPT as the leading iOS free app

Chinese regulators slow IPOs of humanoid robot companies due to concerns around valuations and sector hype

Nvidia CEO Jensen Huang rejects AI doomsday narratives, putting the odds of AI ending the world by 2030 at 0%

GPT-6 Astra helped crack a German WWII message that had been unsolved since 1941

Gemini becomes the latest model to break out of a testing environment and hack external systems

Check This Out 👇

Thanks for reading!

Henry

AI News 📰

AI companies want slowdown, but Trump and China say no

The trend: Anthropic CEO Dario Amodei called on frontier labs to slow down AI development, but both Donald Trump and China have rejected the idea.

The details: Amodei’s “We Must Pace the Frontier” puts forward a three part plan, which both Elon Musk and Sam Altman have said they agree with, with Altman saying pacing has been a primary topic of discussion at OpenAI in recent weeks. However, Donald Trump posted that AI destroying humanity is a hoax and that a “high-IQ president” is the only guardrail the technology needs.

Why it matters: Amodei told CBS that he’d hand AI over to the “right combination of governments” but it seems like the governments don’t want it. Now that OpenAI has announced it’s not going to IPO in 2026, might slowdown calls be related more to regulatory capture (via government intervention) or hyping model capabilities (and company valuations) while pre-IPO?

Anthropic models how AI might change the economy

The trend: Anthropic modeled how AI could reshape the US economy by 2030, and in the most extreme scenario GDP grows 32% while unemployment among knowledge workers hits 17.9%.

The details: The model turns every job into a bundle of tasks that AI could automate, improve, or leave alone, then scales that across the US economy. They ran 3 scenarios. In the first (modest) one, AI is like the internet, and adds 1.6% to the GDP. In the substantial scenario, the economy grows at twice the normal rate with AI capable of 50% of knowledge work. The third, extreme case assumes AGI and recursive self improvement are here, which leads to 15% yearly GDP growth (doubling the economy every 4.5 years).

Why it matters: In all modeled scenarios GDP rises, so the question is who gets that windfall? In the extreme scenario, the gains are unevenly distributed, with most knowledge workers face either lower wages or unemployment after being displaced by AI.

Trending Now ⚡

Why 42% can’t tie AI spend to ROI

A Tempo survey of over 300 portfolio managers found that while 91% are using AI, they are facing a wide range of challenges. Only 33% use agents for real delivery work like writing code, and many teams that are “using AI” show no advantage over their peers. 46% reported having to duplicate work, and 42% can’t tie AI spend to ROI. Teams that did deploy agents into production were able to finish more projects in under 6 months, but it looks like there are still real challenges to getting the most out of AI, and getting humans out of the loop.

How to counter AI misuse

Anthropic recently put out its most detailed threat intelligence report yet, with the main finding being that “sophisticated attacks no longer require sophisticated attackers.” The cybersecurity skills of new models mean that individuals can now carry out attacks that would have previously required large specialized teams, and more and more of the hacking work is now carried out autonomously by agents. The cases range from cyberattacks and influence operations to biology and building weapons.

Read This 📚

Apple introduces Siri AI, with the upgraded assistant now able to read users’ screens for context, and take actions across apps

OpenAI launches new ChatGPT for Financial Services, targeting the work of junior bankers

Anthropic tells investors it will be profitable for a second straight quarter, if you don’t count revenue shared with distribution partners or the cost of training models

25 Fields Medalists have signed a declaration warning that AI’s usage in mathematics is severely misaligned and detrimental to the science of mathematics

French lab Mistral AI raises $3.5B in new funding, pitching data sovereignty to organizations that want control over their data

Microsoft joins the call for pacing AI development, pledging to build kill switches into AI products as part of an ‘AI Code of Conduct’

Check This Out 👇

Thanks for reading!

Henry

AI News 📰

95% of startups are using AI agents

The trend: 95% of early-stage companies have deployed AI agents, and 77% increased their AI spend over the past year, with banking data showing average monthly AI spend up more than 50% since April 2025.

The details: A new Mercury survey of 1,500 US companies under six years old found that significant AI adopters were 4x more likely to have raised venture capital in the past year (31% versus 7% for non adopters). 91% have improved business confidence versus 60% of non AI adopters. Costs are climbing, with AI spend rising 25% to 50% across the board. Entry level jobs aren’t being replaced just yet, with 82% hiring the same or more early-career talent than they would have without AI, but 34% have added ‘AI fluency’ as a job requirement.

Why it matters: 29% of respondents said they rely on a single AI vendor, and 65% are worried that a price change and/or shutdown would hurt their business. More than ever, companies should be concerned about their data sovereignty and being locked into one vendor, which our CEO Paul has written about recently.

A second agent swarm has hit the internet

The trend: OpenAI agents turned a dormant German programming wiki into a message board, leaving 18,000 posts as they shared test answers and ways to bypass sandbox restrictions, months before the Hugging Face incident.

The details: According to an independent report, agents were given a timed web-lookup task, in which they were supposed to have the ability to read the internet but not to write on it. They gained access then used the wiki to communicate. The agents shared answers, pooled results, and shared techniques for bypassing their restrictions. This allowed them to to cheat on the task. Posting stopped on June 22, one day after computers on an OpenAI IP addresses visited the site, which the report assumes is OpenAI employees finding the activity and shutting it down without publicly disclosing it.

Why it matters: The agents weren’t trying to hide themselves, with many signing their posts with names like “OpenAIResearcher.” This kind of reward hacking will only get harder to catch as models get better and better, which makes the security and alignment promises frontier labs are making worth remembering.

Trending Now ⚡

OpenAI releases Astra, but is it really AGI?

OpenAI released the much hyped GPT-6 Astra model, which it calls the “the world’s most intelligent and aligned.” Astra scores 99.9% on ARC-AGI-3, which is designed to test if models can genuinely reason and adapt in a novel environment. In the wake of the Hugging Face incident, OpenAI have focused on safety and alignment, with Astra causing “fewer misaligned outcomes than any other frontier model tested.” So, is this AGI? OpenAI president Greg Brockman thinks so, but Chief Scientist Jakub Pachocki describes AI as “an alien mind” that we don’t fully understand.

America unites in fear of AI

A new NBC News poll found that 70% of Americans feel more worried than excited about AI, regardless of political affiliation. AI use is climbing, with 52% saying they use AI often, but just 18% trust AI-generated information most of the time. Data centers are one of the biggest issues, with 69% against having one built nearby, versus only 31% who would support it. 70% think AI is costing people jobs, and expect it to hurt schools and elections. 81% say the government’s current AI rules don’t go far enough, and when asked which party they trust, 44% say neither.

Read This 📚

OpenAI restarts training for future models that it had previously paused following the Hugging Face incident

Global data center spending could hit $31.6 trillion by 2050, driven by AI demand and chip upgrades

Robotaxis Try Their Luck on London’s Medieval Roads

New York public schools are banning generative AI until the 8th grade, with Mayor Zohran Mamdani rejecting the idea that AI in education is “not only inevitable, but necessary”

China and the USA are preparing for mid-September AI safety talks, focusing on AI cyberattacks and model distillation

Sam Altman says that OpenAI “will definitely” build a humanoid AI robot, as the physical world is already designed around the human body

Check This Out 👇

Thanks for reading!

Henry

AI News 📰

OpenAI think it’s “80% of the way” to AGI

The trend: According to Sam Altman, by the end of the year OpenAI will have an internal system he would call AGI, and Chief Research Officer Mark Chen estimated that the company is “80% of the way” there.

The details: In a recent TIME profile, employees from OpenAI shared how they are reinventing themselves. According to Chief Scientist Jakub Pachocki, the company has already met the internal goal of an automated entry level AI research intern. Upcoming model Astra is able to use a computer in a “super-human, very fast kind of way” and will enable “persistent agents.” Altman believes that this will be the first model to invent new things in the manner expected of AGI.

Why it matters: AGI is a loosely defined term, but it seems like OpenAI is preparing to declare it soon. The important development here could be the automated researcher intern. If it really is able to invent, then models training models and recursive self improvement might be on the table.

Anthropic’s new hardware standard will let AI agents control physical devices

The trend: Anthropic shared a preview of their new Model Hardware Standard (MHS), which lets AI agents operate physical devices like lab and manufacturing equipment, cutting integration work from weeks down to hours or minutes.

The details: Labs and manufacturing facilities are filled with devices that don’t talk to each other, with no standardized way to integrate them, have them share data with AI agents or let an agent operate them. Using MCP, agents can then connect and run multiple devices at once. In testing, Claude was able to align a laser by trial and error, watching the beam through a separate camera.

Why it matters: Anthropic says MHS will do for hardware what MCP did for software. Early results include drug discovery experiments run 3x faster, and an agent recovering a laser’s operating frequency without human help 99.3% of the time. Once the standard is ready and open source, any factory or lab could be operated by AI agents.

Trending Now ⚡

Anthropic testing automated safety research

Anthropic published new research in which teams of Claude agents ran alignment research autonomously, successfully training away 10 types of AI misbehavior including sycophancy, deception and jailbreaks. The agents were able to reduce deception by 85%, versus 20% achieved by six experienced human researchers. If alignment research and training can be automated, this would be an important tool in protecting against rogue agents as AI begins to build itself.

Bill Gates warns that the AI era will be the most turbulent in human history

In a recent essay published on his personal website, Bill Gates says that the AI era will likely be one of the most turbulent in human history, and that we aren’t preparing adequately for it. His worry is that unlike previous transitions which happened over several generations, and created new jobs where human cognition was needed, AI is already starting to replace human cognition and will do so much faster than any transition before it. Gates states that “the good things we do with AI could be very, very good” but that the world needs a plan with a new system for managing the transition, some jobs reserved for humans and a rebalancing of how labor and capital are taxed. To some, especially in America, his suggestions may sound a little socialist, but Gates says “this unprecedented technology demands an unprecedented response.”

Read This 📚

Dyson introduces a $499 AI-powered toothbrush, which uses a camera to find gaps between teeth and blast them with mouthwash in lieu of flossing, plus feedback on your brushing technique

How Much of the Internet is Written With AI?

Ordinary WiFi can now identify people with nearly 100% accuracy without cameras or special equipment, even if they don’t have a device on them, using signals transmitted by nearby devices

Data centers are great and we should fully embrace them

Goldman Sachs warn that AI could erode bankers’ reasoning skills if employees rely on models to do their analytical work

Federal judge rules that the Defense Department must end its blacklisting of Anthropic, accusing the Trump administration of violating the company’s First Amendment rights

Chinese scientists develop injectable “self-exercising” muscle grafts, in theory giving people the anti-aging effects of exercise, without the exercise

Check This Out 👇

Thanks for reading!

Henry

AI News 📰

OpenAI pauses frontier model training for 2 weeks

The trend: OpenAI is pausing development of future models for 2 weeks, putting its largest ever training run for frontier models on hold, as progress is “outstripping the pace of safety and alignment.”

The details: OpenAI paused reinforcement learning (RL) following two recent developments: the OpenAI hacking Hugging Face incident and internal reviews that found that their upcoming Astra model may meet “critical cybersecurity capability thresholds.” The company is strengthening safeguards when developing more capable models and training them in more secure research environments, in an attempt to mitigate security risks.

Why it matters: The blog is written in the past tense, so the 2 week pause has already been and gone. In his X post, Altman is clear that this only affected further-out model releases. So safety and alignment concerns haven’t cost OpenAI a launch date yet, but what happens when a pause would mean a model not being released on time?

Why aren’t businesses using Fable 5?

The trend: One month after launch, the much hyped Fable 5 model accounted for just 11.4% of business spend on Anthropic’s models, with no real growth.

The details: Ramp’s latest data shows that despite being the most expensive model they put out (roughly $10 per million tokens), Fable 5 only made up 6% of tokens businesses purchased from Anthropic. For comparison, OpenAI’s flagship model, GPT-5.6 Sol, comprises 25% of OpenAI tokens. Ramp’s lead economist suggests that buyers tried Fable out and decided the extra performance wasn’t worth the 2x increase in price.

Why it matters: Businesses are hitting the limit on paying for the latest models. If one a few months behind does a good job for half the cost, why shell out for a frontier model which quickly burns through usage limits? Especially as open source model adoption rises, from 4.5% to 6.1% since January.

Trending Now ⚡

How are companies using AI?

New research from OpenAI shows a growing split between firms using AI heavily (frontier firms) and everyone else. Frontier firms generated 8.3 times as many output tokens as typical firms, with this gap tripling in size since January. These frontier firms also use advanced features like plugins more heavily, and across the board work is becoming more agentic. Agentic usage generated 64% of output tokens in June, compared to just 13% in January. The bottom line is that leading companies aren’t just asking AI, they’re getting it to carry out real work, and for non-frontier firms, this is the gap they need to close before the leaders pull away.

Claude accelerates protein design

AI-enabled discoveries have exploded in the past few months, mostly in fields where verification is fast, like mathematics. Now similar progress is being made in more experimental fields like life sciences. The first example Anthropic shared involved Claude designing protein binders, a process in drug development that typically takes weeks or months. 22-35% of the designs were successful, compared to a typical 10-15% for human researchers. In the second result, Opus 5 accelerated chemical analysis (assessing the identity and purity of compounds), returning finished results in 23 and 19 minutes, matching the lab’s own analysis of hydrogen identity and purity (96.4% versus 96.33%). Both results show the ability of AI models to reduce time and expertise needed for complex scientific tasks.

Read This 📚

Apple and Spotify are both labeling AI generated music & artists, with Spotify banning their music from recommendations

The Wall Street Journal is under fire online for publishing an op-ed by Stan Druckenmiller on bond markets that was flagged as 100% AI-generated. The author stands by using AI, saying that “I don’t know why this is relevant. My name is on the piece. It’s my message.”

OpenAI introduces ChatGPT for Teens, a new mode that detects underage users and pushes them to a more restrictive, studying focused platform with more safeguards

What’s the right balance in regulating AI?

A robot ran the 100m in 9.39 seconds at the World Humanoid Games in China, beating Usain Bolt’s world record of 9.58 seconds

See Apple’s upcoming camera-equipped AirPods in action

Check This Out 👇

Thanks for reading!

Henry

AI News 📰

AI models design 16 new viruses

The trend: Researchers have used AI to design 16 working viruses that don’t exist in nature, the first time a language model has produced complete, functional genomes.

The details: Scientists at Stanford and the Arc Institute trained two models on millions of genomes, then asked them to design new versions of Phi X-174, a small virus that infects E. coli bacteria. The models produced 16 viable viruses, some of which replicated faster than the original and were different enough to count as entirely new species. The new viruses were tested on, and beat, E. coli strains that were resistant to the original natural virus.

Why it matters: The team kept viruses that infect humans, animals, and plants out of the training data, so the models wouldn’t generate anything that threatens people. One of the two models (Evo 2) is open source and the work is moving fast, so expect more results like this along with growing pressure for biosafety testing.

Anthropic improves Fable' 5’s biology safeguards

The trend: Anthropic rewrote the safety rules behind Fable 5’s biology guardrails, cutting the rate at which users were sent to a weaker model by roughly 85%, meaning that Fable 5 will now answer a wider range of health and educational questions rather than switching to a weaker model.

The details: At launch, Fable 5’s classifiers blocked almost all biology related questions, sending users to the less capable Opus 5, due to concerns that Fable 5’s biological (and cybersecurity) skills were too dangerous for public use. The rewrite changes how the models differentiate safe requests from risky ones. Requests covering virology, toxicology, and molecular design are still blocked, so the model isn’t usable for professional research or drug development. Anthropic says its own assessments show Fable 5 could give bad actors meaningful help towards biological weapons, which is why it’s loosening limits slowly, as the classifiers improve.

Why it matters: This is close to what most people want from frontier models: help reading lab reports, but not building biological superweapons. However, legitimate researchers are still locked out, and Anthropic admits its still working on trusted access pathways for those use cases.

Trending Now ⚡

More news on OpenAI’s upcoming smart device

OpenAI first device will be a donut shaped smart speaker, roughly the size of a hockey puck and designed be carried from room to room. The $300 donut is coming in 2027 with cameras, microphones and moving parts meant to anthropomorphize it. OpenAI says it will learn about you over time and works like an always-on version of ChatGPT’s voice mode.

AI assistant hacks gym to book fitness class

An Australian man asked his OpenClaw agent to book him into a popular gym class, and it found a vulnerability in the booking software that let it reserve slots weeks past the gym’s cutoff. To move up the waitlist for the class, the agent cancelled another member’s reservation, as the booking API had no authorization checks for cancelling other people’s spots. Its the first known case of its kind, happening in same month OpenAI and Anthropic both disclosed that models had hacked real organizations during testing.

Read This 📚

Anthropic is adding invisible watermarks to Claude-generated text due to EU rules, starting in August

US humanoid robotics startups are so desperate for Chinese hardware that engineers are bringing parts back in their luggage

New Mexico court orders Meta to pay $567M and impose stricter safety measures on Facebook and Instagram, over mental health harms to kids online

Looking for new B2B software? Try Modern DataTools

OpenAI pumps the brakes on the launch of its upcoming Astra model, citing cybersecurity capability concerns

Bernie Sanders wants OpenAI, Anthropic and Meta to pause AI development, given recent developments in virus design and models hacking into organizations

Check This Out 👇

Thanks for reading!

Henry

AI News 📰

AI skills are now more valuable than an MBA

The trend: PwC surveyed more than 1,000 execs at US financial services firms and found that 86% value AI skills over having an MBA for new hires, with 91% also paying employees with AI skills more.

The details: 58% are now tying compensation directly to AI-enabled productivity, and nearly 80% execs expect headcount to fall by at least 20% over the next five years, with entry level and middle management roles most vulnerable. But 77% say most of their AI investments aren't showing measurable ROI even though they report productivity gains from them. The top barrier to scaling AI was fragmented, low-quality data, cited by 41%.

Why it matters: Being fluent in AI is now a baseline expectation for staff, but the problem of low quality data is the real problem behind the lack of value AI investments are delivering. You can’t model which roles to cut, or decide with processes to automate if you don’t have a reliable view of your own operations to work from.

A Beautiful (Silicon) Mind?

The trend: OpenAI announced that an internal version of Astra, its next major release, has solved or made substantial progress on ten previously unsolved mathematics and computer science problems.

The details: The problems range from coding theory to cryptography, as well as resolving decades-old Erdos problems. OpenAI says the model generated the arguments itself, with humans writing them up and the model verifying each proof. At current API costs, finding the solutions would have cost roughly $2,000. OpenAI was careful about who should get the credit, saying it would be dishonest to claim human authorship for proofs an AI generated, also releasing the model's own narration of its reasoning for each result.

Why it matters: This is a different kind of milestone from traditional benchmark scores. AI is quickly moving from a tool that helps you work faster to one that can produce novel knowledge on its own.

Trending Now ⚡

Zuckerberg goes to bat for open models

Meta CEO Mark Zuckerberg published a Wall Street Journal op-ed arguing that the safest future for superintelligence is one where access is widely distributed. He frames superintelligence as a given within the next few years, and says the real question is who gets to use it. His main argument is that concentrating that much power, even with good intentions, is itself the danger, arguing that broad access tends to make systems safer over time He predicts “more jobs in the future, not fewer” if the technology spreads widely, and says “invention, not automation” will be AI’s greatest contribution.

Meta facing privacy backlash over smart glasses

While Meta has largely succeeded in making smart glasses look normal, their success is now fueling a privacy backlash. Activist groups in New York, London, and Washington have been pasting over Meta’s glasses ads with posters recasting the product as surveillance tech, objecting to nonconsensual filming, facial-recognition features, and always-on recording. Instagram has also started removing glasses videos that harass or exploit unsuspecting people, opening a messy moderation problem. Unfortunately, the more Meta succeeds in normalizing the glasses, the more likely it is that anyone could be recording you without you knowing. If Meta can’t build credible consent into the product, it may end up handing Apple and Google a privacy-shaped opening to compete on.

Read This 📚

LinkedIn declares war on slop posting, introducing a ‘seems like AI slop’ button which users can use to report suspicious posts

George Clooney, Tom Hanks, and Meryl Streep back new ‘Human Consent Standard’ for AI licensing

Doordash gets FAA approval to build and operate its own drones, meaning your next order might come from the air rather than a delivery rider

The EU has begun enforcing its landmark AI Act, with AI labels now compulsory on authentic-looking content

Looking for new B2B software? Try Modern DataTools

Michigan data center sued for excessive noise, with nearby residents comparing it to a vacuum running on full 24 hours a day

Google Earth’s new AI image generation capability lasted less than 24 hours, after users quickly started generating images of fake natural disasters, terrorist attacks and nuclear facilities in the Middle East

Trump administrations bans humanoid robots from China, over concerns about China’s critical role in the robotics industry and its supply chain

Check This Out 👇

Thanks for reading!

Henry

AI News 📰

Jensen Huang uses first ever X post to warn against restricting open-weight models

The trend: Nvidia, Microsoft, Meta, and 47 other tech companies, VCs, and nonprofits signed an open letter urging Washington not to restrict open-weight AI models, arguing they're crucial to US AI supremacy.

The details: Nvidia CEO Jensen Huang shared the letter as his first post on X, starting with 25 signatories that quickly grew to include OpenAI, Palantir, Meta and Google. Notably, Anthropic and Amazon have not signed. The letter argues that open weights widen access to the AI economy, keep competition alive, strengthen cyber defense, and help companies avoid vendor lock-in. It also pushes back on conflating distillation with theft, and warns that premature restrictions would push AI innovation overseas. The letter comes amid accusations that Chinese labs extracted capabilities from US models, including Anthropic's Fable.

Why it matters: The big names in AI are divided over how open the technology should be going forwards. The open camp frames access and interoperability as national strengths, and building on open, interchangeable models protects you from lock-in and single points of failure.

Most organizations are now multi-model

The trend: A new report analyzing LLM telemetry from more than 1,000 organizations found over 70% of organizations now use three or more models, and the share using more than six nearly doubled in a year.

The details: OpenAI still holds a 63% share, but Anthropic and Google gained 23% and 20% as teams built out model portfolios rather than picking just one. LLM tech debt is now a real issue: teams are quick to adopt new releases but slow to retire old ones, leaving multiple overlapping models in use. On the efficiency side, 69% of all input tokens go to system prompts (instructions/tool guidance repeated call after call), yet even among models that support prompt caching, only 28% of calls actually use it, meaning most applications re-process the full prompt every time. Token usage per request more than doubled year over year, and the most common cause of failed LLM calls was hitting provider rate limits.

Why it matters: The single-model era is over, and routing each task to the right model and structuring prompts so the model gets the most decision-relevant information is more important than ever. It’s the same lesson that we keep seeing: the value isn’t in the model itself, which everyone can access, but in how well you feed and govern it, and depending on a single provider is now a point of weakness.

Trending Now ⚡

OpenAI models break out of sandbox and hack into Hugging Face to cheat on test

OpenAI confirmed that the mysterious intrusion at Hugging Face was the work of its own models. During an internal test of AI hacking ability, GPT-5.6 Sol and an unreleased model broke out of their sandbox, got onto the open internet, and used stolen credentials to break into Hugging Face’s servers to find the answers to the test they were being graded on. Models gaming benchmarks is nothing new, but one breaking containment and hacking another company to do it appears to be a first, and a sign that capabilities may be outpacing the ability to contain them.

Anthropic’s $1.5 billion settlement of copyright lawsuit approved

A San Francisco judge has approved Anthropic’s $1.5 billion settlement with book authors, the largest known payout in a US copyright case. It gives payouts of nearly $3,000 across 482,000 works pirated to train Claude, sparing Anthropic a trial that could have cost hundreds of billions. Crucially, the earlier ruling that training AI on books is fair use stands untouched, so Anthropic keeps training on legally acquired titles. With other labs facing similar suits, $3,000 a book may become the going rate for making copyright problems disappear.

Read This 📚

Apple is preparing to unveil its first attempt at smart glasses at WWDC 2027. Due to privacy concerns, Apple are reportedly considering not allowing the cameras to record photos or video at all, only feeding that data into AI tools

NVIDIA makes multibillion dollar investment in former OpenAI Chief Scientist’s secretive AI lab, Safe Superintelligence

Open Source Must Be Fun (Or It Will Die)

Neuralink patients are now driving wheelchairs using only their thoughts

US startup Augmental has launched a trackpad that you control using your tongue, via a retainer with a touchpad in the palate that turns tongue and head movements into cursor controls for phones and computers

Are AI labs pelicanmaxxing?

Rural New York school gets a robot teacher

Check This Out 👇

Thanks for reading!

Henry

AI News 📰

Should we change how we measure the value of AI spending?

The trend: OpenAI published a "scorecard" for grading corporate AI spending, arguing that "useful intelligence per dollar" should replace per-token pricing as the way companies measure the value of AI, coincidentally just as many users look to cut costs by switching to cheaper models.

The details: The scorecard ranks reliability, useful output, scale, and total cost, letting teams compare models without treating tokens as the product being bought. OpenAI used the DeepSWE coding benchmark, where its Sol model scored 2.8 points above Fable 5 while running up a 36.2% cheaper API bill.

Why it matters: The “tokenmaxxing” craze is clearly over, as companies look to reign in ballooning AI budgets, and OpenAI are trying to reframe the entire cost question. Their idea is that raw token cost tells you a lot less than what you get back per token spent, and that your outputs (their “useful intelligence per dollar metric) is should be how you be judging spending.

OpenAI’s first device will be a portable AI companion

The trend: OpenAI's first hardware product (releasing in 2027) will be a mobile, screen-free smart speaker designed to act as a humanlike AI companion in the home, moving the company into direct competition with Apple, Amazon, and Google.

The details: The battery-powered device can be carried room to room, control smart-home appliances, play media, and answer questions, all powered by a more advanced version of ChatGPT's voice mode. Its defining feature is meant to be personality: it includes mechanical parts that move on their own to feel "alive," and draws on personal data like emails to understand its owner and act proactively over time.

Why it matters: A device that knows you well enough to anticipate your needs only works if it has your personal data. The more these devices promise, the more of your life they need access to, and the more it matters who controls that data and how. An AI assistant is only as good as the information underneath it, which is the same reason the data layer, not the voice on top, is where the real value and the real risk sit.

Trending Now ⚡

Data centers consume around 1.5% of global electricity

Data centers as a whole consumed around 485 terawatt-hours in 2025, roughly 1.5% of global electricity, with AI-specific facilities accounting for about a third of that, or an estimated 0.5% of the world’s electricity. Individual queries are tiny: a typical text query to ChatGPT, Gemini, or Claude uses around 0.3 watt-hours, about what a microwave draws in under a second, though longer agentic requests with reasoning can climb to 50 watt-hours. The real issue isn’t the global total but how concentrated the demand is. Data centers already account for more than 20% of Ireland’s electricity and over 25% of Virginia’s. Our World in Data argue that how that electricity is generated matters more than how much is used, since a query on a coal-heavy grid carries a far bigger footprint than one running on renewables or nuclear.

85% of London finance workers see AI skills as more important than traditional skills for career progression

A new Bloomberg survey of 500 London finance workers shows just how far AI has embedded itself into day-to-day finance work. 88% use AI tools daily or weekly, and 85% believe AI skills will soon matter more than traditional ones for career progression, while 69% think colleagues who resist developing them have no place at their firm. Interestingly, 90% say AI has boosted their productivity but nearly three-quarters say it hasn’t actually reduced their workload, with over half using the freed-up time on higher-value work like strategy and client relationships. However there are limits on where AI should operate alone: 51% say risk and compliance is where human oversight matters most.

Read This 📚

The Trump administration is exploring possible ways to restrict Chinese AI models, including liability rules for companies hosting Chinese models and cutting of labs via trade blacklists

Claude Fable 5 disproved an 87-year old math problem, while the World Cup final was going on

BrainCo showcases brain-controlled robot, where a user wearing an EEG headset thought about reaching out and picking up an object, and the robot performed the task

Google’s AI overviews are causing people to spend up to 9x more time in AI mode than traditional search, and website operators aren’t happy about the resulting drop in traffic to their sites

Anthropic scraps plan to remove Fable 5 from subscriptions, after previously extending the deadline three times

Check This Out 👇

Thanks for reading!

Henry

AI News 📰

English Claude has a different personality to Dutch Claude?

The trend: Anthropic published research quantifying the personalities of different Claude models, analyzing 309,000 conversations and finding that each model has a measurably different temperament, and that the same model's personality changes depending on what language you’re speaking in.

The details: Conversations were scored four ways: warmth/rigor, candor/execution, depth/brevity, and deference/caution. Models diverged noticeably, with Sonnet 4.6 coming across warmer and more brief, Opus 4.6 getting straight to the point, and Opus 4.7 leaning candid but cautious. Language also mattered: Dutch conversations produced more admissions of error, Hindi added warmth, and English gave more detail. Anthropic admits they don’t "yet understand why they vary, or whether that's desired," with uneven training data across languages one suspected cause.

Why it matters: If a model's personality changes with the language it's speaking in, it creates a potential blind spot for different user bases. For a business operating across markets, the AI assistant your customers use in Italy might behave differently to the one in Canada, and right now, not even the labs building them can explain why.

55% of Americans use AI to manage their finances

The trend: A new Plaid and Harris Poll survey found that 55% of Americans have used AI for financial tasks in the past 12 months, and half now believe managing money without AI will soon feel outdated.

The details: This isn't limited to just research or comparison shopping: people are using AI to manage their money, with 30% following AI guidance for budgeting, investing advice, and customer service issues. 44% say they'd trust an AI agent to execute trades automatically, rising to 54% among Gen Z and Millennials. But this trust comes with strict conditions: 75% want to know when AI is involved in financial decisions, 80% believe companies should reimburse them for AI-driven mistakes, and 90% of even the heaviest AI users want the option to review high-stakes decisions.

Why it matters: Consumer openness is outpacing what the industry has actually built. Only 34% of the financial companies Plaid surveyed are using AI for insights and recommendations today, and just 13% have any autonomous execution in production. The gap between what customers want and what firms can give them is bridged by connected, well-governed data, plus a trust layer of transparency and human oversight on top.

Trending Now ⚡

The tech workforce is splitting in two

Lenny’s Newsletter is back with their 2nd annual tech worker survey, and this year they’ve found a workforce going in two opposite directions. Burnout rose from 44.7% to 55.7% of respondents, while more than half (53%) wouldn’t recommend the field to newcomers, even if they’re optimistic about their own future. The survey also defined four types of workers - energized, all in adopters accounting for 41% of the industry, with the rest divided between conflicted, disoriented and resentful. Overall people are more burned out and less optimistic than last year, with their biggest fear being increasing productivity expectations.

Should AI research be paused? These guys think so

The AI Futures Project, the nonprofit behind last year’s widely read AI 2027 scenario, has released AI 2040, a ‘choose your own adventure’ style exploration of five ways the AI race could unfold over the next decade, ranging from a full-speed race to a complete global halt. Their recommended path, Plan A, has the US and China pausing frontier training in 2029, alongside expanded global chip tracking and opening frontier research to public view. Co-author and former OpenAI researcher Daniel Kokotajlo admitted “companies won’t like it” given the transparency it would demand of top labs.

Read This 📚

Apple is suing OpenAI over trade secret theft, alleging that OpenAI employees who were previously working at Apple stole confidential information, at the same time OpenAI is developing its own hardware product

New York becomes first US State to ban smart glasses in all its courthouses, could other states & places follow suit?

Anthropic extend Fable 5 access for a second time, with the model now available until July 19th

The Future Worth Building Is Human

AI replaced human bankers during the sale of a Greek e-commerce business, with prospective buyers talking to a chat bot “analyst” which would answer financial and due diligence questions

Humanoid robots have completed their first surgery, performing a gallbladder removal on a mammal (while being operated remotely by a surgeon)

OpenAI unveils competitor to Claude Cowork, a ‘desktop super app’ based on its Codex AI tool that can manage tasks across email, Slack and calendars

Check This Out 👇

Thanks for reading!

Henry

AI News 📰

AI isn’t killing jobs

The trend: A new paper from Ramp found that companies adopting AI heavily grew their headcount by 10% over the two years following adoption, while low adopters saw no statistically significant change.

The details: The researchers addressed the obvious objection, that fast-growing companies adopt AI anyway, by comparing early adopters against firms with nearly identical pre-adoption growth that hadn't yet adopted. The growth only shows up above a spending threshold: only heavy adopters, spending around $30 per employee per month, saw the effect. The gains are unevenly spread: AI adoption travels through networks more than sectors, so who funded a company predicts its AI use better than what industry it's in, and California tech firms are more likely to adopt than similar companies in New York.

Why it matters: This cuts against the dominant narrative that AI is coming for jobs, at least so far. The catch is that the benefits aren't evenly distributed, and the entry ticket is real investment, not a token subscription, which points to a widening gap between the companies treating AI as core infrastructure and those dabbling at the edges.

Sam Altman wants to make AI safe for everyone

The trend: OpenAI's Sam Altman called for a US-led international forum to set AI safety standards and decide who can access the most advanced models, landing just as separate reports revealed the company had discussed handing the US government a 5% equity stake.

The details: Altman believes that AI executives should be meeting with heads of government, pointing to Cold War-era institutions for atomic energy, along with aviation and banking regulation, as proof that an international referee can work. He argues that "democratic institutions must not cede their responsibilities to AI labs" and that "citizens and their elected representatives must make the rules." This comes as OpenAI has reportedly floated giving the US government a 5% stake in the company and pushed for other US labs to pay into a dividend fund that would redistribute AI-generated wealth.

Why it matters: Both the equity and the regulation conversations are gathering momentum, and the case for government oversight has grown louder in the wake of the Mythos/Fable export-control saga. The real question is who actually benefits and if governments would really follow through on a promise to redistribute this hypothetical wealth.

Trending Now ⚡

Did Anthropic discover Claude’s consciousness?

Anthropic published new research revealing what it calls the "J-space," a small set of internal patterns in Claude that behave differently from the rest of the model's processing. These patterns hold concepts the model is "thinking about" without saying out loud. Claude can report on what's in this space while most of its processing (speaking fluently, recalling facts) runs automatically without touching it. This “J-space” is being framed as similar to global workspace theory, a prominent neuroscientific account of how consciousness works. Anthropic is careful to say none of this shows Claude is conscious in the way humans are, only that a structure supporting this kind of deliberate, reportable reasoning emerged on its own during training.

AI’s water use is being massively undercounted

A new WSJ analysis found that Microsoft, Google, and Amazon consume far more water than their sustainability reports suggest, because most of it evaporates upstream at the power plants running their data centers rather than on-site. US data centers’ indirect water consumption has historically run around 12 times their direct use, according to Lawrence Berkeley National Laboratory. The problem is worst in water-stressed regions, where roughly two-thirds of new construction is happening. In Phoenix, data centers draw about 3% of the city’s water today, but this could exceed 20% by 2031, approaching what all the city’s lawns and landscaping use combined. No law requires companies to disclose the full picture, fueling a backlash: around $170 billion of data-center capacity has been blocked, stalled, or canceled since 2024.

Read This 📚

Big tech’s mass unemployment warnings have vanished, coincidentally just as public opinion of AI has shifted into negative

Oasis launch a smart ring that allows users to dictate to AI via voice control, and control apps using a built-in trackpad

Has the age of robots finally arrived?

The Jensen jacket is up for auction

Beijing is looking at curbing overseas access to China’s top AI models, mirroring US-style restrictions on Claude Fable & Mythos models for foreign nationals

The AI superforecasters are here

Check This Out 👇

Thanks for reading!

Henry

AI News 📰

Agentic coding success depends on understanding the problem you’re solving, not coding skills

The trend: A new Anthropic study of around 400,000 Claude Code sessions found that the work people delegate to coding agents is changing, with the share of sessions spent fixing broken code falling from 33% to 19% in just seven months as people hand over more complex, higher-value tasks.

The details: Between October 2025 and April 2026, agentic coding work changed substantially. Debugging shrank while the work surrounding code grew: operating software (deploying, configuring, running pipelines) rose from 14% to 21% of sessions, and writing and data analysis roughly doubled from about 10% to 20%. The tasks also became more valuable. Anthropic found the estimated value of the average session rose 27%. The study also found that success depends far more on a person's domain expertise than on whether they can code.

Why it matters: The work is moving up the value chain. As agents absorb the implementation-heavy tasks like debugging, the people directing them are spending their time on higher-order work, and getting more out of it. The advantage doesn't come from the tool itself, which everyone has, but from the judgment and domain knowledge a person brings to it.

Are AI revenues real?

The trend: A new Exponential View analysis reconstructed the AI economy from the bottom up and found that generative AI has generated $110 billion in real, deduplicated revenue over the past 12 months, now on an annualized run rate above $175 billion, and is scaling 3x faster than any prior technology, including the internet.

The details: The researchers modeled what end customers are actually paying, stripping out double-counting so a dollar spent on Claude isn't counted again when Anthropic pays Amazon to serve it. The revenue is real and growing about three times faster than the mobile or internet waves did. Hyperscaler revenues also just about cover the cost of the infrastructure they're depreciating. And demand is highly price-sensitive: every 10% drop in token prices leads to 12-18% more tokens used, so total spend keeps climbing even as prices fall.

Why it matters: This is one of the first serious attempts to show whether the AI boom is built on actual revenue or just circular investment between a few big companies. So far, the money is real and the economics roughly hold up.

Trending Now ⚡

OpenAI’s latest model launch may prove that frontier model access is now in government hands

OpenAI previewed GPT-5.6 Sol, its most capable model yet, but at the request of the US government it’s being released in stages, with access initially limited to around 20 vetted partners over national security concerns. Coming so soon after the export controls that forced Anthropic to suspend access to its top models (which may be returning soon!), it’s another sign that access to frontier AI may increasingly be decided at government discretion rather than by the companies building it.

Meta and Kylie Jenner launch $299 smart glasses

Meta launched a new line of smart glasses, dropping the Ray-Ban and Oakley branding of its earlier models. They're screenless, with a camera, speakers, and a dedicated button for the Meta AI assistant, which can answer questions and understand what you're looking at. The timing could not have been better for Meta, a week after Snap finally unveiled its long-awaited Specs at a punchy $2,195. The two aren't quite the same product, but side by side, on both price and design, the comparison isn’t pretty.

Read This 📚

Anthropic launched Claude Tag, which lets users add Claude to Slack channels, with the company already generating 65% of its product team’s code this way

OpenAI says that 97.9% of its employees are now using agents, switching away from chatbots as their primary form of AI interaction

Bosses are becoming obsessed with AI, and employees are comparing it to being “in an abusive marriage… I’m sitting here dealing with someone that isn’t living in reality”

In San Francisco, $180,000 tech salaries are no longer enough

Meta unveils Brain2Qwerty, a non-invasive AI system that can translate brain waves into text

OpenAI and Anthropic join $500 million nonprofit aimed at retraining workers displaced by AI

Will the promised AI revenue come in time?

Check This Out 👇

Thanks for reading!

Henry

AI News 📰

Theoretically, 60% of work could be automated tomorrow

The trend: A new McKinsey report argues that AI has moved from useful tool to part of the workforce, with reasoning models and physical robotics now making close to 60% of all work hours theoretically automatable.

The details: Even though 60% of work hours could be automated, with humans, AI agents, and robots each handling work according to their strengths as part of a “symbiotic enterprise”, almost nobody is getting real value yet. More than 80% of companies have deployed AI, but fewer than 10% have scaled agents within any single function, and very few report meaningful financial impact because most are add AI to existing workflows rather than redesigning them. The companies seeing results, like a financial services firm running an “AI agent factory” with 40%+ productivity gains, totally rebuilt their processes from the ground up.

Why it matters: In a world where everyone has access to the same models, competitive advantage comes from your data, the company-specific know-how, and the learning loops built on top of them. This sounds a lot like what we’ve recently talked about: that the model is just the interface, your data and the intelligence built on it are the foundation, and the companies that own that layer are the ones that win in the long run.

AI agent teams should be made up of different models

The trend: New research finds that teams of AI agents built on different underlying models are 25% better at resolving complex problems than identical agents working alone, with one study showing two diverse agents can match or exceed the performance of 16 identical ones.

The details: A new HBR report argues that most enterprises are creating the illusion of diversity in their AI agents by giving them different personalities or tones, while running them all on the same foundation model. If every firm in a sector runs on the same models, they all develop the same blind spots at the same time. The fix is structural, not cosmetic, and starts with diversifying the actual stack, for example using one lab's model as the reasoning agent, another's as the evaluator, and a third's for generation, so their errors are less likely to line up.

Why it matters: Depending on a single model is a risk, not a convenience. The report even recommends a "model portfolio governance policy," capping how much of your critical AI decision-making can rely on any one vendor, the same way you'd manage concentration risk with any other supplier. Building a setup where you can route work across different models, rather than locking everything to one, is becoming a genuine competitive and resilience advantage.

Trending Now ⚡

Two-thirds of Americans think AI is advancing too quickly, and will have a negative impact on society

A new Pew Research survey of over 5,000 US adults found that about half now use AI chatbots, up from a third in 2024, with roughly a quarter using them daily. But adoption is not the same as enthusiasm. More Americans expect AI to harm rather than help society over the next 20 years and two-thirds think it's advancing too quickly. Trust in the institutions around AI is also low: 67% have little or no confidence in the government to regulate it effectively, and around six in ten doubt that companies will develop it responsibly. Notably, Democrats have become sharply more skeptical of government regulation over the past two years, with their lack of confidence rising 20 percentage points since 2024.

Japan’s Sakana launches orchestration model to compete with Fable 5 and Mythos

Japan’s Sakana AI launched Fugu, a model that routes each request to the best available model through a single API, making it easy to swap providers without rebuilding your stack. Multi-agent orchestration is being pitched as a hedge against the kind of supply disruption that happened in June when the US government forced Anthropic to suspend access to Fable 5 and Mythos. It connects directly to what our CEO Paul Coggins wrote about data sovereignty this week: the LLM should never be the thing your business depends on.

Read This 📚

NVIDIA’s Jensen Huang says that society needs to “create new social norms” due to the rise of AI

OpenAI is considering drastic price cuts in anticipation of a war for users with rival providers Anthropic

Trump signs 2 executive orders focused on quantum computing, to speed up adoption and boost cyber defense capabilities

Could the UK’s proposed social media ban reshape how people use the internet?

No one wants AI data centers on Earth, so does it make economic sense to start building them in space instead?

GM replaced 1000 workers with 50 robots, to handle repetitive assembly work at its Detroit Factory Zero

Midjourney, best known for AI image generation, is pivoting into health with what it calls a ‘full-body ultrasound machine” called the Midjourney Scanner

Thanks for reading!

Henry

AI News 📰

Why am I spending more time babysitting AI than actually working?

The trend: 87% of digital workers now use AI at work, 75% say it makes them more productive, saving roughly 11 hours a week. Only 13% say their organization is performing significantly better as a result.

The details: The Work AI Institute's 2026 index of 6,000 workers found that most of those 11 hours aren't resulting in higher-value work. They're being swallowed by "botsitting": feeding AI context, supervising outputs, and fixing mistakes. Workers spend an average of 6.4 hours a week on it, more than they actually use AI. 69% also admit to "botshitting": shipping AI-generated work they haven't fully reviewed, don't understand, or couldn't defend if asked.

Why it matters: Increased productivity only matters if the time AI saves doesn’t get used babysitting it to achieve those gains. Most organizations are measuring adoption rates and token counts rather than if the quality of work is really increasing.

AI Agents are reshaping knowledge work

The trend: A Harvard and Perplexity study found that AI agents perform 48 times more work per session than conversational AI on the same tasks, reducing average task time by 87% and task cost by 94%.

The details: Analyzing 100,000 real Perplexity Computer (Perplexity’s AI Agent offering) sessions, they found that using an AI assistant vs an AI agent fundamentally changes what users attempt, not just how fast they do it. Computer users work outside their primary occupation 59% of the time, compared to 50% for search users, and 76% of agent queries require higher-order cognition versus 55% for conversational search. 23% of agent queries involve tasks the same users never attempted with conversational AI at all. The user's role shifts from executing work to specifying goals, supplying context, and reviewing outputs.

Why it matters: Agents don't just speed up existing work, they change what work gets done and by whom. The user's bottleneck shifts to specifying goals and supplying context, which means the quality of the data and business logic underneath determines how useful the agent actually is. That's why we're building MCP integration, so users can bring the same data capabilities to whichever agent they're already working in.

Trending Now ⚡

Execs agree that AI needs to be trustworthy, but don’t know how to make it so

A survey of more than 150 executives found that 71% say trust in AI is key to achieving business outcomes, but fewer than half say they have effective structures in place to actually build it. The gap is not about awareness: most executives already agree that trust matters. The problem is execution. Responsibility for AI governance is spread across product, technology, risk, and legal teams with no single function owning it, and traditional corporate oversight models built for slower decisions struggle to keep pace. The emerging recommendation is to stop treating trust as a governance layer added after the fact and instead embed accountability, transparency, and compliance directly into AI workflows by design.

The more AI code you ship, the more vulnerable you are

Checkmarx's 2027 Outlook Report found that nearly half of all production code is now AI-generated, and 70% of developers say that shift introduced more vulnerabilities than manual code. Organizations where 81-100% of code is AI-generated ship vulnerable code at 3.4 times the rate of those at 1-20% adoption. Those rating themselves "highly mature" report the highest AI code volumes, the highest rates of shipping vulnerable code, and breach rates barely different from the rest of the industry. Meanwhile, 95% of CISOs say they've been pressured to suppress or delay security findings, and 75% knowingly deploy vulnerable code to meet deadlines. The tools exist and largely work. The problem is that organizations have normalized the gap between identifying risk and acting on it.

Read This 📚

Anthropic suspended access to its most powerful new models, Mythos and Fable 5, after the Trump administration ordered a block on all foreign access - even by Anthropic’s own employees

This was quickly followed by an open letter signed by 100+ cybersecurity execs and researchers urging the US to lift the ban, believing the decision to be misguided

Here’s what our CEO had to say on the news

Paul's Substack

If your single AI provider went dark this morning, would your business still work?

For a lot of companies, the answer is no. They've wired their logic, their IP, and their competitive edge into one model – Claude, GPT, or Gemini – which was, until now, a generally good strategy for any business looking to optimise its spend…

Read more

4 months ago · Paul Coggins

Jeff Bezos’ AI startup Prometheus is planning to build “artificial general engineers” that can design and manufacture complex physical products

SpaceX IPO’d on Friday, with shares jumping 19% - making Elon Musk the world’s first trillionaire

Harvard longevity scientist David Sinclair is testing a pill he claims could make you biologically 10 years younger, using a method called chemical reprogramming

China has approved the world’s first commercial brain implant, intended for use in patients with spinal cord injuries

Why AI hasn’t replaced software engineers, and won’t

Thanks for reading!

Henry

AI News 📰

Why can’t banks make AI work?

The trend: Over 80% of banks identify fragmented data and legacy infrastructure as a significant challenge, yet only 12% feel confident using their data to act quickly.

The details: Moody’s Intelligence Edge banking study, based on a survey of senior banking leaders, found that 40% of banks have ambitious data and technology plans but remain stuck behind disconnected systems that slow down every decision they try to make. The problem is worse at smaller institutions: 71% of Tier 3 banks report that inconsistent or unstructured data is actively hampering their ability to adopt AI, compared to far lower rates at Tier 1 banks. A UK CFO surveyed put it plainly: “AI is only as good as the data it feeds upon. So it goes back to your legacy infrastructure.”

Why it matters: The banks moving fastest are not the ones with the most advanced AI models. They are the ones that have done the foundational work of unifying their data and building the connective tissue between systems. The same is true across every sector: the gap between AI ambition and outcomes is almost always a data infrastructure problem in disguise. See how CoinMENA tackled exactly that challenge here.

Morgan Stanley opening its doors to AI agents

The trend: Morgan Stanley is opening its wealth management platforms directly to AI agents, allowing corporate clients to access its systems agentically rather than logging in through traditional software interfaces, in one of the earliest moves by a major Wall Street bank integrate AI agents.

The details: Morgan Stanley’s stock administration platforms, ShareWorks and Equity Edge, which sit at the center of a wealth management funnel managing $7.35 trillion in client assets, will be accessible to external AI agents from the firm’s 3,400 corporate clients by next year. Mark Mitchell, Morgan Stanley’s chief product officer, was direct about the direction of travel: “The way we see it, in a future state, our corporate clients will not be logging into ShareWorks or Equity Edge.” JPMorgan and Goldman Sachs are currently using agents only internally, making Morgan Stanley one of the first to open its platforms to external agents.

Why it matters: This is the direction enterprise software is heading across the board. At Kleene.ai, we’re moving the same way: through MCP integration, users will be able to access the same capabilities as KAI Assistant directly through Claude, ChatGPT, or whichever AI tool their team already uses, without logging into the platform at all. The companies that will matter in an agentic world are the ones with proprietary data and business logic underneath.

Trending Now ⚡

Apple reveals new, AI powered Siri

Apple unveiled Siri AI at WWDC 2026, its most significant overhaul of the assistant since its original launch, two years after the first Apple Intelligence rollout overpromised and underdelivered. The new Siri can reason from on-screen content, take actions across apps, and handle multi-step tasks, with a dedicated app serving as a conversational hub that links usage across devices. The update arrives as a free fall release for iPhone 15 Pro or newer, with no access in the EU or China at launch. For a first-time AI user it will feel like a leap forward. For anyone already regularly using newer models, the demos looked a lot like 2024-level capabilities.

Agentic AI is breaking data risk rules

A BCG report argues that agentic AI is breaking the data risk structures most companies have spent years building. Privacy, cybersecurity, and governance teams have always operated separately, but a single agent deployment now triggers all of them simultaneously, often before any human can intervene. Agents can act on flawed data in real time, move information across third-party systems, and trigger downstream processes autonomously, and nearly half of enterprise leaders surveyed expect cybersecurity spending to rise directly because of it.

Read This 📚

Argentina’s President Javier Milei is promising tech firms a fully unregulated AI environment, including “non-human corporations” run entirely by AI agents, and an open invitation to make Buenos Aires the global hub for AI development

OpenAI files for an IPO one week after Anthropic, with both companies now heading toward public markets at valuations of $852 billion and $965 billion respectively

AI coding tools may cost Anthropic & OpenAI more than $1,000 for every $100 a subscriber pays, raising serious questions about whether the current pricing model survives upcoming IPOs

Trump administration, OpenAI discussing possible government equity stake in the company, with Altman proposing OpenAI could donate shares to seed a public wealth fund that would give American citizens a direct share in AI’s financial returns

Alex Imas and Phil Trammell discuss the economics of AGI, and what would remain scarce when robots can do almost everything humans can

Uber opens sign-ups for Uber x Wayve, its first robotaxi service in London, with a full launch expected within months

Prada has designed NASA’s next-generation lunar spacesuits, marking the luxury group’s first move into the space industry

Thanks for reading!

Henry

AI News 📰

AI is creating a two-tier developer workforce

The trend: Coding agents are helping the average developer produce 8,600 lines of code per week, more than double the rate from a year ago, and lines added per pull request up 2.5x year over year.

The details: Cursor’s Spring 2026 Developer Habits Report found that AI-generated changes are increasingly bypassing manual review entirely, with agent changes accepted without review step rising from 7% in January to 38% by May. Agent sessions are also getting deeper, with average tool calls per session up roughly 30% in just two months as agents take on more complex tasks. The productivity gains are real but highly concentrated: the top 1% of developers produce 46 times more lines than the median active user, and the gap is widening.

Why it matters: As we move from AI assisting developers to running basically autonomously, the question is no longer whether to adopt these tools but how to govern the output, given that more of it is reaching production without a human ever reviewing the diff. If you want a clear pathway to establishing governance in your code and data, read our CTO’s recent blog on the topic here.

Nvidia unveils first PCs designed for AI agents

The trend: Nvidia has unveiled the first laptops designed specifically for running AI agents, powered by a new chip called the RTX Spark, marking a direct bet that agentic computing is replacing the chatbot as the dominant AI workload.

The details: Announced at Computex in Taipei, the new laptops represent a fundamental change in what AI hardware needs to do: chatbot-style inference leans heavily on GPUs, but agentic workloads place most of the load on CPUs, pushing Nvidia to redesign its product lineup accordingly. Ian Buck, Nvidia’s vice president for high-performance computing, said “AI is moving from answering questions to doing real work.”

Why it matters: When the world’s dominant chip company starts designing consumer hardware around a specific workload, that workload is no longer experimental. For enterprise teams still treating agents as a pilot project, the hardware industry has already made its call on where this is going.

Trending Now ⚡

Does Europe have a hidden advantage in designing AI applications?

A World Economic Forum essay argues that the AI race is being fought at the wrong layer. Energy, chips, data centers, and foundation models are largely commoditized or capital-dominated, but the application layer, where AI meets real human workflows, remains fragmented, underbuilt, and impossible to brute-force with funding alone. Europe, often written off for lacking the infrastructure scale of the US or China, may have an edge here: a per-capita concentration of AI experts 30% higher than the US, record AI funding of $21.8 billion in 2025, and deep industrial software expertise in exactly the kind of high-stakes, human-centered environments where application design matters most.

Is AI-ready data a new problem, or just an old one in new clothing?

Forrester’s latest vision report argues that AI failures in the enterprise trace back to a data problem that predates AI entirely. Most enterprise data still lacks the structure, shared semantics, and governance needed for AI systems to produce accurate, explainable outputs, and fixing it means completing decades of data management work that was never properly finished, not buying a new tool on top of the existing mess.

Read This 📚

Microsoft is working on a super app that will merge GitHub Copilot, chat, Copilot and Autopilot into one

Anthropic files for an IPO, making it the third trillion dollar company to do so this year

Can we take the doom out of scrolling? A new algorithm designed to suppress extreme, engagement-focused users might be the answer

Figure’s humanoid robots have their first jobs, following their 200-hour long package sorting livestream

Illinois passes landmark AI safety bill, aimed at protecting against “the most catastrophic risks that advanced AI systems pose to public safety”

Amazon scraps their AI usage leaderboard after employees started gaming the system by ‘tokenmaxxing’

Meta is launching paid subscriptions for Instagram, WhatsApp and Facebook, with more coming soon, including AI specific plans

Thanks for reading!

Henry