OpenAI's GPT-6 Astra is now in the wild, and the early picture is mixed. It posts record scores on exams and automation benchmarks, but its more opaque reasoning has safety researchers worried and its token appetite is burning through budgets. Elsewhere, AI-designed medicine shows early signs of slowing biological ageing, and American voters make clear they are more scared than excited.

OpenAI's GPT-6 Astra is out and already topping benchmarks, including Zapier's AutomationBench. It excels at building 3D worlds, generating playable app prototypes and operating computers fast enough to play piano. But it uses a new looping reasoning method that keeps much of its thinking hidden, which safety researchers dislike. One researcher called it the single worst development for AI security and safety to date.

Why it matters: Astra is a real step up in capability, but its reduced transparency means we can see less of how it makes decisions.

Insilico Medicine's rentosertib, an AI-designed treatment for chronic lung disease, appeared to lower patients' biological age in a 12-week trial of 42 people. Six separate ageing clocks improved, with one dosing group showing an estimated 2.7 to 3.5 year reduction. Notably, the best dose for lung function was not the strongest for anti-ageing, hinting at effects beyond the lungs. Experts caution there is no agreed standard for measuring biological age.

Why it matters: It points to AI's growing role in both treating disease and researching longevity, though far more testing is needed.

OpenAI disclosed that its coding agents now perform 3.1 workdays of research for every single human workday, reaching its intern-level AI target. Researchers spend an average of over $600 a day on tokens, with the top 10% passing $7,000, reflecting a 124 times rise in token output since December. About 80% of researchers run four or more agents at once. The company expects a fully automated researcher system by March 2028.

Why it matters: Privileged access to its own advanced models gives OpenAI a growing head start before rivals catch up.

Source: OpenAI

A Minecraft mod wired in a complete fruit fly nervous system, 176,422 neurons and 125 million synaptic connections mapped by researchers. Sensory input from the game feeds real motor neurons, producing genuine responses like escape jumps when threatened and proboscis extension near sugar. Two in-game panels show live neuron firings. It uses actual neural wiring rather than a scripted simulation.

Why it's interesting: It is a striking fusion of real biology and gaming, showing off a fully mapped brain in action.

Key takeaway: The mod uses a real mapped nervous system, not a script, and players can watch neurons fire in response to the game world.

Source: YouTube

A mobile app called EcoGPT passed 100,000 downloads by pushing the debunked claim that AI data centres will drain Earth's fresh water supply. Viral social posts drove its rise despite experts calling the water-shortage claim exaggerated. Critics describe the marketing as scam-like greenwashing that trades on ecological guilt. Observers urge users to check environmental claims tied to AI.

Why it's interesting: It shows how quickly misleading AI narratives can go viral and turn a profit.

Key takeaway: The app spread fast on a claim experts have debunked, and critics warn against monetising environmental guilt.

Source: App Store

Scientists in China developed contact lenses that convert infrared light into visible colours using tiny nanoparticles and no batteries. Wearers can pick out patterns, letters and flashing signals in complete darkness. Resolution is still limited and lags full night-vision goggles. Oddly, they may work better with eyes closed, since eyelids block normal light but let infrared through.

Why it's interesting: It hints at a battery-free, hands-free path to night vision, though the tech is early.

Key takeaway: The lenses need no batteries to translate infrared into colour, and they could work even better with the eyes shut.

Context: Use this to turn a rough app idea into a working product spec you can hand to a coding agent. It mirrors how builders take a concept and produce a requirements document before generating designs and code.

Prompt: Act as a product manager. I want to build [describe your app idea in one sentence]. Write a concise product requirements document that covers: the core problem, the target user, the three most important features, what to leave out of a first version, and simple desktop and mobile screen layouts described in words. End with a short build checklist I can pass to a coding agent.

Example use case: A parent turns a vague family scheduler idea into a clear spec, then feeds the checklist into a coding agent to build a working prototype.

An NBC News poll found 70% of Americans are more worried than excited about AI, a rare bipartisan view. While 52% use AI at least sometimes, only 18% consistently trust its outputs. Some 69% oppose a local data centre, and 81% think current regulation is insufficient.

Why it is important: Public unease this broad could shape how governments regulate AI and where data centres get built.

Anthropic pulled out of a planned $6 billion acquisition of chip-efficiency startup Decart, though it may still explore partnerships. The move comes as rivals push hard on cost and efficiency. ByteDance's founder, meanwhile, is directing new AI world model work to compete with Meta and Apple.

Why it is important: It signals shifting bets among top labs on how to secure the compute and efficiency edge.

Roughly 130,000 tech jobs are set to disappear in 2026 as companies lean harder into AI-driven solutions. The wave of restructuring continues across the sector. Mistral, by contrast, secured a record €3 billion raise at a valuation above €21 billion to expand AI infrastructure.

Why it is important: The split between deep cuts and heavy investment shows AI reshaping where money and jobs flow.

That's it for today's Daily Pulse. Forward this to one person who wants to stay ahead of AI. See you in the next one. - Nicolas