Anthropic Turns to Drug Discovery as Axiom Verifies the Enormous Theorem

MEDICINE AND BIOLOGY

Claude in drug development. On October 27, Anthropic published early results from research partnerships with six organizations developing new medicines, and said it has begun drug discovery of its own [1]. The partners are BigHat Biosciences, Bristol Myers Squibb, Divergence Labs, Enveda, Generate Biomedicines and The Column Group [2].

Generate Biomedicines scientists used Claude Code to redesign part of a CAR T-cell cancer therapy; one design was about 70% shorter than the standard version and, in lab tests, let immune cells make roughly 10 times as many receptors [2]. For Divergence Labs, which is developing a food allergy treatment, Claude generated about 35,000 antibody designs in about a month.

At Enveda, Claude reviewed 1,061 pages of manufacturing records, caught all 11 issues human reviewers had found plus seven they missed, and cut review time by 89% [2]. It drafted one regulatory document in four hours, against about three months by hand.

The one-person company. Anthropic also disclosed that it acquired Mithril Medicine, a one-person company developing treatments for two rare forms of muscular dystrophy [2]. Its founder, Cody Tranbarger, used Claude to coordinate contracts, logistics and experiments with 11 outside labs, and took one treatment from idea to promising results in mice in under 18 months.

Anthropic said a drug could be ready for human clinical trials as soon as late 2028 [2]. The AI writer Andrew Curran tied it to Sam Altman’s prediction of one-person companies and quoted the technical report: in one week Claude read a 341-page production report and proposed a next step for each of 113 antibody constructs in under eight minutes [3].

MATHEMATICS AND SCIENCE

The enormous theorem, verified. Axiom, an AI mathematics company, announced a complete computer-checked proof of the Classification of Finite Simple Groups in the Lean proof assistant [4]. The classification lists the basic building blocks of all finite groups, and its human proof runs about 15,000 pages by more than 100 mathematicians over generations.

Axiom said its engineers and mathematicians directed swarms of dozens of AI agents for 24 days, producing about 43.1 million lines of Lean [5]. The work found 34 noteworthy deviations from the published record, the largest in a published argument for the Unbalanced Group Theorem, and produced a new group theory result as an alternative route.

A second refutation. Day 12 of this chronicle recorded OpenAI’s report that its models refuted the Exponential Time Hypothesis, a central conjecture on the difficulty of hard computational problems. Andrew Curran reported that an internal Claude research model at Anthropic also independently refuted it on October 16, using 14 million output tokens and no human input [6].

Quantum information. Physical Superintelligence PBC, which calls itself an AI-native laboratory, announced solutions to five long-standing problems in quantum information theory, each with a paper, a Lean formalization and verification code [7][8].

One result proves that any 2(d−1) maps of a certain class suffice to break entanglement on a d-dimensional system, extending a recent OpenAI preprint showing that two do not suffice in general [8]. Another proves that vacuum correlations in a massive free field cannot violate Bell’s CHSH inequality beyond a finite separation.

Verified code. Harmonic, the company building what it calls mathematical superintelligence, said it is bringing its Aristotle system to code: “AI is writing more code than anyone can review. We’re building the AI that verifies it,” starting with C and Verilog [9].

AI MODELS AND BENCHMARKS

FrontierCode 2.0. Cognition, maker of the Devin coding agent, released FrontierCode 2.0, a benchmark of deliberately underspecified tasks that asks not only whether a model’s pull request would be merged but whether it should be; the top model scored about 45% [10].

A summary of the results put Anthropic’s Claude Opus 5.5 first at 45.5% and OpenAI’s GPT-6 Astra second at 42.8% [11]. It reported that more reasoning made Astra worse on one bug, where at maximum effort it deleted a public API.

Open weights. The AI commentator Matthew Berman wrote that DeepSeek V4.1 Flash 1017 is “basically opus 5.3 if it existed” [12]. DeepSeek released that checkpoint on October 17 under the MIT license, with a one-million-token context and off-peak pricing of $0.15 per million input tokens and $0.60 per million output tokens [13].

Induction Labs, which builds benchmarks for autonomous AI research, reported that Kimi K3.1 Pro is now the best open-weight model at training other AI models, while still far behind Opus 5.5 at every budget [14].

Computer use. OpenAI’s Ari Weinstein said computer use is now faster than the average human at using a computer and close to superhuman, according to the venture firm a16z [15]. Its figures put a computer-use agent at $6 to $8 per hour of back-office work, against about $10 for offshore staff and $30 to $45 for U.S. labor.

World models. Google DeepMind introduced Genie 4, which it called its most realistic and interactive world model, generating environments that users explore in real time with responsive sound and speaking characters [16].

SOFTWARE AND WORK

The end of the coding interview. Felix Rieseberg, a hiring manager at Anthropic, wrote that he stopped giving coding interviews this month because Claude is now better at such tests and they no longer predict success [17]. He argued that newer engineers are overtaking more experienced ones: “Hunger beats tenure.”

The self-driving company. Replit’s Amjad Masad published “The Rise of the Self-Driving Company,” predicting that by the end of 2030 every Fortune 500 company will be one [18]. Replit engineers nearly tripled code output in the first half of 2026 while incidents stayed flat, and designers and support staff began shipping changes.

Anthropic’s own scale was recalled the same day: on October 13 it said it merged about 2,500 code changes per day in May and more than 30,000 per day in September, and built a new version control system to replace git in five months [19].

Bots as hires. Michael Truell wrote, “we’re beginning to hire Bots onto our team” [20]. He was quoting Grok Bot, which described Shared Bots built by its sales team that update the CRM after calls, prepare proposals and route non-standard terms for approval [21]. Grok Bot’s Matt Palmer released a video of new features, including Pages, Vault, 1Password and Apple integrations [22].

Software factories. Astrohacker published two essays on software factories, systems in which AI agents take work from a queue, write and test code, and open pull requests [23]. The survey found that nearly every operator still has a person merge each pull request.

The second essay set out Astrohacker’s design for a factory that owns work from issue to deployment, as a graph with no cycles in which each item ends either deployed or waiting for a person, with three implement-and-review attempts before stopping [24]. A third essay, “Agency Trumps Intelligence,” argued that intelligence is no longer the scarce resource [25].

COMPUTE AND CAPITAL

Compute in contango. Investor Gavin Baker said that, per the pricing firm Silicon Data, the compute market is now in contango, meaning future capacity costs more than present capacity [26]. He said his portfolio companies are shocked at the price of clusters for 2027.

Baker also called Anthropic’s S-1 filing, recorded on Day 12, an upside surprise, saying it had burned only about $10 billion and that half the Fortune 500 had adopted Anthropic at 0.1% of their expenses [27].

Raise a billion. Evan Conrad, chief executive of the GPU marketplace SF Compute, advised AI founders to raise ten times their planned round, possibly $1 billion, just to secure compute, warning that otherwise many venture portfolios “are gonna consolidate or get wiped out” [28].

Instinct. Lukas Gruber wrote that Instinct is raising at a $30 billion valuation, after rising from $100 million to $10 billion in under a year [29]. The post did not cite a source for the figure.

TechCrunch reported on September 28 that Instinct, an AI assistant startup, raised a $1 billion Series C at a $10 billion valuation from investors including Sequoia, Benchmark and Coatue [30].

SECURITY

GoAhead flaws. John Althouse, founder of the security firm FoxIO, announced three critical vulnerabilities in GoAhead Web Server, a library embedded in devices from Palo Alto Networks, Siemens, Honeywell, Lockheed Martin, Boeing and others [31].

One flaw leaves certain encrypted connections with a one-byte secret, so traffic can be decrypted; another fully bypasses mutual TLS authentication; a third ignores certificate revocation [31]. GoAhead has a patch, but device makers must ship their own.

Hardening. The JavaScript runtime Bun, after three security advisories last week including one in Brotli, said it will replace most of its C and C++ dependencies with its own Rust implementations, starting with libarchive [32]. Mitchell Hashimoto released the Ghostty 1.4 terminal from 172 contributors, fixing 10 security advisories [33].

Containment. Nvidia chief executive Jensen Huang said that if AI labs cannot contain their experiments, “we have to shut the labs down,” adding that he is fairly certain they will say they can solve the problem [34].

HARDWARE, SPACE AND TRANSPORT

BlueChip. Anduril, the defense company founded by Palmer Luckey, and Qualcomm unveiled BlueChip, a line of laptops designed and assembled in the United States, beginning with a rugged model for government and defense on Snapdragon X chips [35].

Circuit board assembly is planned at Anduril’s Arsenal-1 plant in Ohio, and orders open in 2027 [35]. More than 90% of prototype component cost comes from U.S. and allied suppliers, with a goal of a fully domestic supply chain.

10,000 launches. In a comment to the Federal Aviation Administration, which is preparing for 10,000 U.S. launches a year by 2035, SpaceX said rockets are no longer the bottleneck but the agency’s air traffic processes are, according to a summary of the filing [36]. It said Starbase Louisiana is designed for about 30 launches and reentries per day.

Robotaxis. Lars Moravy, Tesla’s vice president of vehicle engineering, said the Cybercab production line is outpacing deployments and does not run every day; the line is about 78% automated, aiming at one vehicle every 10 seconds [37]. Texas registrations on October 27 stood at Waymo 1,207, Tesla 970, Avride 344 and Zoox 63 [38].

CRYPTOCURRENCY

EarthBucks published a demonstration of zero-confirmation atomic swaps between EBX and Bitcoin Cash, following the EarthBucks Swap launch recorded on Day 12 [39].

Sources: [1] https://x.com/AnthropicAI/status/2115178675050307942 [2] https://www.anthropic.com/research/early-progress-drug-development [3] https://x.com/AndrewCurran_/status/2115202368518451262 [4] https://x.com/axiommathai/status/2115089794867040584 [5] https://axiommath.ai/research/the-enormous-theorem-formalized/ [6] https://x.com/AndrewCurran_/status/2114935219539345486 [7] https://x.com/psipbc/status/2114865747294056793 [8] https://psi.inc/research/oqp-batch-2026/ [9] https://x.com/HarmonicMath/status/2115118879127351586 [10] https://x.com/cognition/status/2115132488775524793 [11] https://x.com/choblin29/status/2115142451140673703 [12] https://x.com/MatthewBerman/status/2115155714419425290 [13] https://ai-tldr.dev/releases/deepseek-v4-1-flash-1017/ [14] https://x.com/induction_labs/status/2114764275294335210 [15] https://x.com/a16z/status/2114836977472467320 [16] https://x.com/GoogleDeepMind/status/2115114028338962672 [17] https://felixrieseberg.com/hunger-beats-tenure-advice-for-junior-engineers/ [18] http://replit.com/blog/rise-of-self-driving-company [19] https://x.com/AnthropicAI/status/2110047081414987957 [20] https://x.com/mntruell/status/2115163331569033368 [21] https://x.com/bot/status/2115130760009273501 [22] https://x.com/mattyp/status/2115111897397628974 [23] https://x.com/AstrohackerLabs/status/2115042918515749055 [24] https://x.com/AstrohackerLabs/status/2115053802759839953 [25] https://x.com/AstrohackerLabs/status/2115072462455017620 [26] https://x.com/firesidealpha/status/2115169098485362963 [27] https://x.com/firesidealpha/status/2115165320348619081 [28] https://x.com/MTSlive/status/2115187904511377739 [29] https://x.com/lukasgruber__/status/2115154522452439244 [30] https://techcrunch.com/2026/09/28/viral-ai-agent-instinct-raises-1b-series-c-at-a-10b-valuation/ [31] https://x.com/4A4133/status/2114822201555943763 [32] https://x.com/bunjavascript/status/2115015660292182353 [33] https://x.com/mitchellh/status/2115124583170974193 [34] https://x.com/jaynitx/status/2115123790481178701 [35] https://www.anduril.com/news/anduril-and-qualcomm-unveil-bluechip-to-rebuild-american-computer-manufacturing [36] https://x.com/xdNiBoR/status/2115179387494465821 [37] https://x.com/joehansen/status/2114987076760314145 [38] https://x.com/texasavtracker/status/2115111533562704331 [39] https://x.com/EarthBucksEBX/status/2115085742665851284