• EVO 2 was trained on 9 trillion base pairs; its developers reported performance competitive with or better than supervised human-specific models on a cancer-risk-gene mutation benchmark across coding and non-coding variants. They flagged distal-enhancer prediction as a weakness, which another group improved through a different data composition. Y Combinator
• EVO 2’s few-shot prompting generated novel landmodulin candidates from 14 known examples and CRISPR-Cas9 systems from four examples. In a separate design task, an enhancer produced with prompting after reinforcement-learning post-training was reported active and specific in the mouse brain. Y Combinator
• Virtual-cell work faces a data-diversity bottleneck: agents expanded an NIH transcriptomics dataset from 36 million to 500 million-plus cells, but immune cells—especially T cells—were overrepresented, prompting a shift toward AI-guided experimental data generation. Y Combinator The Locus system also improved over off-the-shelf coding agents on molecular-binding and generative tasks as GPU compute scaled, but those results were in silico and still require lab validation. Y Combinator
• AMA co-founder Alfredo described building a personalized research lab for each cancer patient using AI agents to analyze patient data; he cited a rare brain-tumor case where cross-tumor analysis identified potential therapies despite limited relevant literature. Y Combinator Clinical translation remains a constraint: clinicians retain treatment authority, setting up a new NF1 trial is a separate challenge, and hospital tissue handling can prevent organoid work. Y Combinator

• Hassabis described Isomorphic Labs as a new spinout building on AlphaFold technology to reimagine drug discovery with AI; he framed reducing drug development from an average of 10 years and billions of dollars to months or possibly weeks as an ambition. Sony AK
• AlphaFold predictions for 200 million known proteins were made freely available without restrictions; Hassabis said more than two million researchers use the resource and that it has become a standard biology research tool. Sony AK
• Hassabis’s screen for applying game-derived AI to real-world problems is whether a task has a large combinatorial search space, a clear objective to optimize, and sufficient data or experience—or an efficient, accurate simulator. Sony AK
• DeepMind’s stated next step is to combine agent-based search and planning with Gemini and world models for real-world tasks; Hassabis forecast major robotics advances in the next two to three years. Genie2’s generated game worlds, however, currently remain consistent for only a few seconds, with longer consistency an active development goal. Sony AK

Garry Tan says his gstack-evals project tried to reproduce an article about using Jev with Kimi K3, and reports that it replicated; he says Particle-1 did “even better.” @garrytan +1 He gives no evaluation metric or numerical results. @garrytan

• Google says it aims to compete at the frontier while using its full stack—from hardware and models to products—to serve users across a range of cost and quality levels and use cases. Lenny's Podcast
• Google product leader Tulsi described a “60/20/20” approach to AI products: make roughly 60% work now, build another portion around capabilities that seem near, and reserve room for more ambitious features as models improve. She cautioned that rigid wrappers can constrain newer models, making repeated retesting important. Lenny's Podcast
• For AI product evaluation, she recommended starting with small, hands-on tests—around 5–20 prompts—rather than beginning with a large benchmark, and evaluating the end-to-end experience because judge quality, serving, filters, and mitigations can all affect results. Lenny's Podcast
• Google’s stated safety approach includes monitoring models for possible misalignment and intervening; the interview also cited work with Wiz to identify and patch vulnerabilities, including one found in a hospital system. Lenny's Podcast

• Garry Tan argues Apple should retool its software team around “software factories,” saying teams of a few people with 100 code-generation agents could remake buggy apps such as TV and Calendar. @garrytan
• He cites TV app defects: downloading resets the list, and trying to move a movie to the top of the download queue crashed the app and erased the queue; he says the reset is easy to identify in telemetry and fix. @garrytan
• Tan says frontier models make software easier to change, making bugs less tolerable, and predicts personal just-in-time software will become inevitable. @garrytan

Garry Tan urges founders to build their own “software factory” @garrytan, amplifying Demis Hassabis’s forecast that one person who understands AI could outperform an entire startup team and that AI-enabled builders could run whole companies with one-tenth the headcount in five years @RoundtableSpace.

Martin Casado questions METR’s report for not normalizing AI progress by spending, asking whether the analysis is useful without that adjustment @martin_casado. The comparison in the quoted post says AECI progress per $1 billion of R&D fell from 4.05 points in 2025 to 1.49 in 2026—a 63% decline—even as progress per year doubled @michael_wiebe.

Garry Tan says he is conducting open-source evaluations of harnesses with frontier models and links a GitHub project called GHarnessEvals, describing the work as being done in public. @garrytan +1

Garry Tan said he was asking his “software factory” to test a prompt for Claude Code or Codex that asks agents to identify implicit state machines in an app and improve user, developer, and agent experience without breaking anything; he said he would report back, so no test result was included. @garrytan +1

Y Combinator’s Paper Club framed AI biology as shifting from reading biological data toward modeling and designing living systems: sequencing costs have collapsed, data is outpacing researchers’ ability to interpret it, and models are beginning to catch up. @ycombinator The session covered cancer genomics, virtual cells and autonomous research, DNA language models, and few-shot prompting for biological design. @ycombinator

Paul Graham says YC fosters unusually close founder relationships: 20 years ago, many YC founders lived in a San Francisco building nicknamed the Y Scraper, and a reunion prompted his claim that YC founders become closer friends than founders backed by any other investor. @paulg

Martin Casado said he was unsure whether Seb Krier’s graph was correct, but that what it showed was what he would expect from prior technology trends; he called anticipating a different outcome this time an aberration warranting scrutiny. The post does not specify the graph’s subject or findings in text. @martin_casado

A podcast summary says Nikesh Arora grew Palo Alto Networks’ market value from $18 billion to more than $250 billion; after eight years and 35 acquisitions, the company had more than 70 products. @davidsenra The described acquisition playbook is to buy winning teams, put the acquired founder in charge, agree on a product roadmap before signing, and target only number-one or number-two players. @davidsenra The promotional post discloses paid partnerships. @davidsenra

David Sacks attributed Republicans’ midterm losses to the war and its economic effects, while saying job and GDP growth remained strong and inflation would drop quickly if the war ended @DavidSacks. He cited Brent crude’s 10% election-day decline as a market reaction to the possibility of an earlier end to the war, and argued an end would rapidly reduce energy and transportation costs @DavidSacks.

Sarah Guo says AI still cannot help with the iterative parts of writing an essay: ordering ideas, cutting excess length, maintaining coherence, and preserving substance. @saranormous

Jason framed U.S.–China AI competition as a national-security race, calling AI the battlefield and saying Jensen Huang and NVIDIA’s technology were critical targets that needed protection, while urging the U.S. to accelerate toward superintelligence. @Jason The attached MaggieNYT post reported that a highly classified memo overseen by Marco Rubio said the intelligence community assessed that a woman who had formed a close relationship with Huang was working for China’s Ministry of State Security. @maggieNYT

For evaluating efficiency-driven markets, Martin Casado suggests looking beyond price: efficiency can lower prices and create a rebound, so compare aggregate margin dollars before and after the efficiency gain. @martin_casado

A shared post says an open-world game was made using Opus 5.5, gpt-Astra, Three.js, Anatomy, and Suno, with an art direction inspired by Moebius and Roger Dean; Martin Casado praised it as beautiful. @bei351009016 +1

Scott Kupor says Sequoia has handled generational and management transitions unusually well over its 50-plus-year history, a rare capability that has helped it remain a tough, respected competitor among venture firms.@skupor

• WashU’s Scott Wilson says the endowment’s AUM grew from $8B to $19B in nine years. It held SpaceX at more than 20% of its portfolio and used continuation vehicles to stay invested through the IPO; Wilson says the approach depended on having the right governance structure. @HarryStebbings
• On portfolio construction, Wilson says diversification benefits taper beyond 20–30 look-through exposures; WashU targets half its portfolio in its top 50 positions. He sees megafunds as a poor fit unless a round is too large for WashU to invest directly, and doubts they can repeat the multiples of smaller funds. @HarryStebbings
• Wilson believes Anthropic could have raised privately at more than $2T, a valuation he doubts public markets would support, and attributes weak DPI to capital chasing private-company champions. WashU also wants valuation discipline and avoids putting capital into deals at 100× revenue. @HarryStebbings
• WashU identifies direct opportunities by asking managers which five portfolio companies excite them most, then meeting promising companies and offering to help lead the next round; it has also led employee tender offers and invested directly rather than waiting for SPVs. Wilson names not adding to winners—partly because they were already 1–2% of the portfolio—as his biggest mistake. @HarryStebbings