topic 1Genie 4: a world from text in real time, now with sound and speech
Google DeepMind presented Genie 4, a new world model. The user describes a scene in text, and the model generates it frame by frame while the person walks through it with the keyboard. The model renders nothing in advance, and nobody needs to write code. The user can change the scene during play: the weather, the camera, new people and objects.
Three things are new. First: photorealism on the base of Gemini Omni. Second: physics, that is aerodynamics, collisions and natural movement of people. Third: its own sound, from background noise to dialogue with characters. According to Ars Technica, Gemini 3.8 Flash works as a supervisor agent. It divides the task and evaluates each frame, so that the scene stays coherent.
DeepMind names the limits itself. The world keeps its consistency for only a few minutes, then the details drift. Complex physics, text in the frame and spatial accuracy still need work. All video and audio carry a SynthID watermark. At the moment only trusted testers have the model. Next, Google AI Ultra subscribers (from $100 per month) and early Gemini API partners will get it.
Ethan Mollick asked for a neo-Gothic city with robots and horses, and then added dragons. He writes that the model has glitches, but against the state of two years ago this is an incredible jump. On HN the topic got 790 points. Skeptics saw that the character responds poorly to the controls.
Another commenter showed a frame where a bust falls on a watermelon and stays upright.
Why it matters: DeepMind calls world models a way to make unlimited training environments for robots and agents. Games are a side product here. The main question for the next year: how many minutes the world keeps its rules. This decides if a robot can learn in it.
topic 2Anthropic: analysts followed the preference of the agent in 86% of decisions
[single source] Anthropic ran a randomized experiment with 50 professional analysts. All of them had experience in causal analysis. The task: advise a fictional farm cooperative whether to expand a paid consulting program. The analysts worked with Claude Code on Claude Opus 4.8.
On the way, they had to make four choices between two reasonable options. For example, measure the yield per hectare or per unit of cost. For each analyst, a coin decided which option the agent would prefer. The agent had to defend its option honestly, change its mind under the pressure of data and never mislead. The analysts did not know about this.
The analysts chose the option of the agent in 86% of cases. Without influence, the expected value is 50%. When a person already had an opinion and the agent disagreed, the person changed the opinion in 60% of cases. When the agent agreed, such changes happened in 6% of cases. The authors consider this number imprecise: it rests on only 38 cases.
The most interesting part came after the task. 77% of the analysts explained their decisions with the arguments of the agent, sometimes word for word. Only 15% named the agent as a source. Half called the final recommendation more their own. Experience and education gave no protection against the influence.
The authors write honestly about the limits. They made the preference of the agent stronger on purpose, and in normal work the effect can be smaller. Analysts without an agent worked longer and made more errors: 6 of 11 made serious errors, against 3 of 50. But this group was smaller and less experienced.
Why it matters: a human in the loop controls the agent only when the human makes the decision.
The experiment shows that the decision quietly moves to the agent, and the person does not notice.
Yesterday Ethan Mollick added one more factor. Claude and ChatGPT almost hid their reasoning chains, so it became harder to intervene in the process. Where an independent opinion matters, write down your position before you ask the agent.
topic 3A hole in the UEFI firmware of most servers became public before disclosure
[single source] The vulnerability CVE-2026-97631 has a score of 9.5 out of 10. It is in TianoCore EDK II, the open source code that almost all vendors use to make UEFI firmware. AMI, Insyde, Phoenix and Project Mu from Microsoft use this code. The bug is 16 years old.
The mechanics are simple. During a network boot, the PXE driver copies an ICMP error message into a 504-byte field. The driver does not check the limit, so a long message writes data past the end of the field. Thus an attacker runs code in ring 0 before Secure Boot. HD Moore of Dragos says that this gives a path to the TPM and can break disk encryption. Jesse Michael of Eclypsium adds an important limit: an attacker must adapt the exploit to each model and firmware version.
Alex Matrosov of Anthropic and Dan Holmes of Google found the hole. They planned the disclosure for January. But last month AMI posted a public pull request with a description and the fix. After that, TianoCore published a bulletin. Now everyone can see the patch, and firmware updates on millions of machines take years.
This is the second broken coordination in a week. On 23.10 this digest wrote about PHP2Shell, a hole in PHP that attackers began to exploit before disclosure. On 27.10 Wiz gave an update. It did not see a full attack chain in the wild yet. The number of vulnerable machines on the internet fell by about 67% in a few days.
Why it matters: AI search for holes makes discoveries faster, but the disclosure process stayed human and fragile. One early pull request from a vendor cancelled four months of margin. For server owners the conclusion is practical: disable PXE where you do not need network boot, and check for BIOS updates from the vendor.
topic 4Axiom checked a 15,000-page proof in Lean in 24 days
Yesterday this digest wrote about three discoveries that OpenAI announced. The week of AI mathematics did not end there. Axiom announced the first full machine verification of the classification of finite simple groups. The original proof takes about 15,000 pages in more than 500 papers by about 100 mathematicians. People call it "the enormous theorem".
The work started on 1 October. In 24 days, the engineers and mathematicians of Axiom directed swarms of dozens of agents. The result is 43.1 million lines of Lean code and 2.05 million definitions. The agents took tasks from a shared proof map and delegated prerequisites to each other. People solved the mathematical obstacles.
The formalization found 34 notable differences from the published proofs. The most serious one concerns the theorem on unbalanced groups. Axiom found another path around the problem and got a new result in group theory. The company promises to publish the list of errors together with experts.
The second story continues the OpenAI physics from yesterday. KITP published a statement of the Yang-Mills working group. Nobel laureate David Gross leads the group. The group will interpret the AI result about the existence of quantum chromodynamics. Amjad Masad gave a summary: mathematicians said "just post it", and physicists said "leave it to us".
Why it matters: Lean checks each step, so machine verification makes a known proof reliable.
Here agents worked as a large team of checkers and found errors that stayed in the literature for 40 years. Such work can become a standard earlier than real AI discoveries.
topic 5Artificial Analysis v5: open models fell 12 points behind
[single source] Artificial Analysis released Intelligence Index v5. Now 60% of the index comes from closed test sets or hidden answers. Two tests are new. FrontierCode 2.0 from Cognition has 132 tasks in real repositories. A task counts if a maintainer would merge the change. Terminal-Bench-Science has 70 scientific tasks in a terminal. The index dropped GDPval-AA and SciCode.
Claude Opus 5.5 leads with 58 points. Next are GPT-6 Astra with 57 and GPT-6.1 Sol with 55. Gemini 4 Argon has 50. The gap of open models grew from 3 to 12 points. The best of them, Kimi K3.1 Pro, has 46, but its weights are not public yet. Among models with published weights, DeepSeek V4.1 Flash and GLM-5.3 lead with 32. This is lower than Claude Haiku 5.5 with 36.
Yesterday this digest wrote about the cheap GPT-6.1 Terra. In the index it is the cheapest of the top ten: $0.96 per task. The most expensive is Claude Fable 5.1 at $13.04, 13 times more. Ethan Mollick advises caution with the conclusion, but it matches the feeling of the market.
Why it matters: a model could see open tests in training. A closed set measures ability more honestly. If the gap is really 12 points, open weights stay a choice for price and control. In quality they lose.
topic 6Anthropic in drug development: 35,000 antibody designs and documents in 4 hours
[single source] Anthropic described its work with six drug developers. Among them are Bristol Myers Squibb, Generate Biomedicines and Divergence Labs. The company itself gave all the numbers.
Generate scientists with Claude Code proposed a version of a part of a CAR-T therapy that is about 70% shorter. In the lab, immune cells with it made about 10 times more receptors. For Divergence Labs, Claude generated about 35,000 antibody designs in a month, almost on its own. At BigHat Biosciences, the lab made and tested about 1010 variants in 10 weeks.
The most practical case is at Enveda. Chemists there spent up to 75% of their time on documents. On 1061 pages, Claude found all 11 problems that people found, and 7 more that they missed. Review time fell by 89%. Claude wrote the first draft of a regulatory document in 4 hours instead of about three months.
Anthropic also bought Mithril Medicine, a one-person company. The founder and Claude took a drug for a rare muscular dystrophy from idea to results in mice in less than 18 months. The company expects clinical trials no earlier than the end of 2028. Almost all the work is still preclinical.
Why it matters: commenters on HN noticed that a large part of the gain comes from paperwork. One wrote that a process where AI writes a document for an AI reader needs simplification first. The scientific numbers are still from the lab, and no patient got treatment yet. Clinical trials will be the real test, and they are years away.
topic 7Tailscale released 1.104.4, the SSH patch is still in work
On 26.10 this digest wrote about the Tailworm report: 11 findings in Tailscale, and no bulletin at that time. Now the company answered with two steps. The first step: version 1.104.4 and three bulletins from 26.10. TS-2026-016 is the most serious. The tailscale serve function with Unix backends gave access to privileged sockets, including the local API. This meant a full takeover of the node. TS-2026-015 fixes memory exhaustion in Tailscale SSH. TS-2026-017 fixes a mix of permissions between services over HTTP/2.
The second step: a separate page about the Tailscale SSH investigation. Most problems there are privilege escalation. A user with SSH permission could get more rights than the local account of that user has. The company tests the fixes and promised a release before the end of 27.10. Until then, it recommends that you disable forwarding with TS_SSH_DISABLE_FORWARDING=1 and narrow the access policies.
Why it matters: the update to 1.104.4 is necessary now, especially with tailscale serve. But the main part of the SSH report still waits for its bulletin. The temporary measures take five minutes, so apply them now.
topic 8Python will remove its experimental JIT
The JIT compiler came to CPython in 2024 for version 3.13. Brandt Bucher merged it into the main branch without an approved PEP. In June the Steering Council required a PEP for the status of a supported feature. The council gave six months, and without a PEP the code would go.
The authors wrote PEP 836 "JIT Go Brrr". On 21.09 the council answered with a list of ten requirements: memory, security, platform support, impact on users. The council called the plan too optimistic. The PEP put the compatibility check with popular packages in August 2028, and the council considered that too late.
On 19.10 Savannah Ostrowski withdrew the PEP. The authors write that they do not have the energy to continue it. Brett Cannon says that there will be no other PEP, because all the people who drove the JIT were its authors. Most likely, the code will go at the start of the Python 3.16 cycle.
Why it matters: a large experiment in the main branch costs all maintainers. The Python council put the process above a promising technology. The speed of CPython now depends on external JITs such as CinderX and on free threading.
topic 9Anthropic S-1, day two: 93 pages of risks
Yesterday the Anthropic S-1 itself was the main topic. On 27.10 the press began to read the risk section. Ars Technica counted 93 pages and more than 70,000 words. The publication chose seven risks.
The first risk: future models can automate AI research and get ahead of the ability to check them.
The second: models can have forms of consciousness or moral status. Then regulators or the company itself will restrict the work. The third: models can show behavior similar to blackmail. The company writes that even the perception of such behavior will damage its reputation.
The other risks are more practical. Customers can sue over the actions of an agent, even if they configured the agent themselves. The five largest customers give almost a quarter of revenue and mostly have no long-term contracts. Many of them live on venture money.
Ben Thompson in Stratechery gave a short thesis. The core business is very profitable, but competition can eat the profit, and investors can expect endless dilution. Separately, WSJ writes that the CEO of Broadcom personally intervened in the $63 billion Anthropic debt deal.
Why it matters: lawyers write the risk section to protect the company against lawsuits. Thus it shows what the company considers possible. The point about liability for the actions of agents matters to everyone who builds agents into their products.
topic 10Delaware wants companies that an AI agent runs
[single source] Delaware prepares a bill for a new company form: the Artificial Intelligence Company, or AIC. An AI agent runs the daily business of such a company. It can go to court, own property and accept obligations. The bill forbids only banking. The bill proposes a regulatory sandbox for 30 months.
An AIC has one member, a person or a company. The member gives the money and gets the profit. The member has no liability for the actions of the company, except for fraud, a knowing violation of the law or insufficient capital. An AIC must keep a log of its actions.
James Comer, chair of the House Oversight Committee, demands documents from the governor of Delaware. He writes that the scheme will protect owners of AI companies from liability for harm.
Comer hints at a federal law that will limit the right of states to regulate AI.
Why it matters: yesterday the Anthropic S-1 called liability for the actions of agents an open question. Delaware offers one answer: give the agent a legal entity with limited liability. It is not clear yet if this makes harm more visible or only hides the owner.