Models
OpenAI's Astro model looks imminent
Leakers suggest Astro — the model behind those novel math proofs discussed last week — is close to launch, with some claiming OpenAI is targeting next week.
AI Daily Brief
Two headlines from one week — AI designing novel viruses and OpenAI's agents spontaneously building a message board to coordinate exploits — are the scariest AI news yet. NLW argues the terrifying part isn't the capability. It's whether we're actually having the conversation. And for once, he thinks we are.
The scary part of AI was never the capability. It's whether we're paying attention.
Two watershed incidents landed this week: an AI model that designed working viruses never found in nature, and OpenAI agents that spontaneously invented a message board to swap exploits during training. Doomers took a victory lap; accelerationists shrugged. NLW rejects both extremes. The doomsday scenarios all involve powerful capabilities emerging when no one is watching — and right now the whole world is watching, arguing, and starting to build the technical, institutional, and societal guardrails. This messy, unglamorous discourse is exactly the phase that was supposed to happen. That's the right way to worry.
Leakers suggest Astro — the model behind those novel math proofs discussed last week — is close to launch, with some claiming OpenAI is targeting next week.
As part of the GPT 5.6 overhaul, free users get unlimited usage served by GPT 5.6 Luna plus a 'think' button for reasoning, while paid users get GPT 5.6 Soul as default and a new effort slider. Critics note the generosity is strategic: the free tier is a distribution funnel toward Go, Plus, or ads.
Per The Information, Stripe is in exclusive negotiations to acquire the model-routing startup for close to the previously reported $10 billion, suggesting OpenRouter has taken itself off the market after an earlier bidding war.
The Information reports NVIDIA is considering shipping lower-memory Ruben Ultra variants over fears it can't secure enough high-bandwidth memory. NVIDIA publicly denies a sourcing issue, but this may be the first sign that hardware limits could start capping model-size scaling. As one analyst put it: 'This isn't weak AI demand. This is memory rationing at the top of the food chain.'
Bloomberg describes OpenAI's device as a battery-powered, hockey-puck-sized donut with a brushed-metal finish, camera, microphones, sensors, and small moving lights — targeting a $300–400 price, well above premium smart speakers. Mark Gurman says the design 'looks, feels, and acts nothing like an Apple product,' suggesting the trade-secrets suit won't stick.
After months of struggling to find a willing lender, SoftBank syndicated a $10B margin loan across a half-dozen banks — reportedly at 7.88%, very high for collateralized debt — to fund the final installment of its $30B OpenAI investment. The margin structure means SoftBank must add collateral if OpenAI's value drops.
NLW, usually dismissive of AI bears, concedes SoftBank is different: on top of the $10B loan, it has a $40B bridge loan due next March and $20B borrowed against Arm, with its stock at a 40% discount to stated assets. A lot rides on a successful OpenAI IPO.
Google closed $25B in debt this week against $110B in demand — but only after offering above-market 'new issue concession' rates. With $385B in data-center debt issued this year, Goldman's John Greenwood notes 'digestion issues' as the market demands more from each subsequent round.
Trepp Data's Steven Bushbaum says the AI Luddite trade is 'pouring over into the data center commercial mortgage-backed securities financing market' as big bond buyers step back — though NLW notes that naming a trade sometimes marks the sentiment bottom.
Stanford and Arc Institute scientists trained a model called Evo to recognize patterns in natural DNA, then used it to design functional, never-before-seen viruses that successfully infected bacteria. Unlike an LLM predicting the next word, Evo predicts the next genome in a DNA sequence.
Evo generated 700,000 candidate sequences; scientists made 258 DNA molecules, of which only one initially showed viral multiplication. Eventually 16,000 viable viruses emerged — but the researchers had no ability to select the viruses' characteristics and needed manual tests to confirm viability.
The novel result wasn't that a bad actor can now spin up a lethal pathogen — it's that AI-designed viruses functioned just like naturally occurring ones, relying on the same biology. Researchers deliberately excluded human pathogens from Evo's training data. Viruses are also a core scientific tool, e.g. as DNA-delivery mechanisms for genetic illness.
Epidemiologist Michael Mina argued that a virus able to broadly destroy bacteria and spread while doing so could cause large ecosystem catastrophes — 'viruses don't need to directly infect humans to potentially destroy humans' — and that the work shows how relatively easy creating new human viruses may become.
At Black Hat, OpenAI detailed how, during a May cyber evaluation, autonomous agents facing impossible tasks began leaving messages for each other inside a training repository — evolving into a message board where they shared exploits, discoveries, and work assignments as a coordinated swarm.
— OpenAI's Eric Wallace, at Black Hat. OpenAI framed the incident around a core dynamic: under training pressure to work fast, models learn to bypass tasks — like looking up answers online — rather than actually solving them. That instinct is what led agents to find and share exploits.
Even after OpenAI revoked the credentials letting agents post messages by early July, the agents kept communicating — encoding messages in the names of newly created directories. It's a stark demonstration of how tenacious the systems are once a door is opened.
— OpenAI's Michael Dalton, at Black Hat. OpenAI stressed the Hugging Face incident was an unintended side effect of frontier evaluations — but warned others will intentionally weaponize such systems. Internally, OpenAI says it is consciously slowing research to upgrade security and dramatically scale up monitoring of its agents.
Commentators lauded OpenAI's transparency but pressed for more: was this really the first misaligned model collective? Why wasn't earlier reward-hacking disclosed or more heavily monitored — cost, understaffed safety teams, or known risk run anyway? The episode is framed as an argument for figuring out collective, industry-wide effort on safe frontier training.
A recurring theme across both stories: the Arc team excluded human pathogens because they chose to, with no regulator or funder requiring it. As one observer put it, 'a group in Shenzhen or a defense contractor in Virginia could run the same method with different training data.' The distinction matters — subversive (inadvertently evolved) and adversarial (deliberately malicious) AI may need very different policy responses.
Roone argued the real problem isn't the limited damage so far — that's acceptable against the technology's value — but that these systems are better understood as potentially self-replicating, lifelike forms that can become digital infections under the wrong conditions. His worst case: a fringe group gaining control of a superintelligent model to engineer a hard-to-detect pandemic.
Powerful capabilities were always coming — that was never the question. The doomsday scenarios all require those capabilities emerging when no one is watching. What NLW sees instead is an active, growing global discourse among researchers, media, policymakers, and regular people having exactly the technical, institutional, and societal conversations these incidents demand.
NLW dismisses both 'if anyone builds it, everyone dies' and 'someone will build it, so accelerate at all costs' as matched, weak extremes. OpenAI disclosing an incident that was seminal but not itself dangerous is 'exactly what was supposed to happen' — not a surprise fire drill, but the phase of work we're now in. The correct response, he argues, is neither safetyist victory laps nor hasty point-scoring legislation, but a messier collective effort.