Teahose.
SIGN IN
NEW HERE — WHAT TEAHOSE DOES
We read the entire AI & tech firehose — so you don't have to.
PODPodcastsAll-In, No Priors, Acquired…
NEWNewslettersStratechery, Newcomer…
PAPPapersPhysical AI research
PHProduct Huntdaily launches
VCInvestor ScoutSequoia, a16z, Benchmark…
CLAUDE DISTILLS →
7 reads, 30 sec each — free, 6 AM ET.
+ a live graph of the companies, people & themes underneath.
HOME/晚点聊 LATETALK/189: Geodesic顾全全:预训练有点“无聊”了,但用sc…
POD
// EPISODE
晚点聊 LATETALK

189: Geodesic顾全全:预训练有点“无聊”了,但用scaling改变AI制药才刚开始

DATE November 19, 2026SOURCE 晚点聊 LATETALKPARTICIPANTS 程曼祺, 顾全全
// KEY TAKEAWAYS6 ITEMS
  1. 01Scaling is the general-purpose lever, and AI-for-science has barely been scaled
  2. 02Compute is the moat, so AI drug discovery is now a capital-intensive foundation-model game
  3. 03Vertical integration: owning the wet lab is the second leg of the stool
  4. 04A triggering event: Chai/Trillion-style antibody design (Try2) showed the technology had crossed a threshold
  5. 05Fast strategic pivot inside a big company: six months from signal to beating the benchmark
  6. 06Pretraining is an organizational capability, not a secret
In this episode

1. Key Themes

Scaling is the general-purpose lever, and AI-for-science has barely been scaled

Gu Quanquan's core thesis is that scaling is not specific to LLMs. He argues the AI-for-science field never scaled because its practitioners are mostly not large-model people and lack the compute. He describes SeedFold (900M parameters) as only a first-step validation: "SeedFold and SeedPro2 can also be seen as the first step of scaling verification, because doubling model size and doubling data size already worked" [00:23:19]. He states the generality directly: "scaling is the most effective method I've seen or used in my twelve years" [01:29:39]. On why others haven't done it: "most people in AI for Science didn't come out of large-model training; they may have good data intuition, but their knowledge and experience in scaling is limited" [00:23:47].

Compute is the moat, so AI drug discovery is now a capital-intensive foundation-model game

Gu's pitch is that AI-bio without heavy compute is just another biotech. He says: "if your thing doesn't need compute and a few hundred cards can do it, this company in my view has no competitiveness... anyone can raise a bit of money, buy a few hundred cards, and worst case poach someone" [00:57:10]. He frames the investor reaction: "when I tell them my compute alone is $10 million in six months, they're shocked" [01:06:21]. He also notes that the lab buildout is cheaper than compute: "a few million dollars" [01:05:52].

Vertical integration: owning the wet lab is the second leg of the stool

Gu argues post-training needs experimental reward signals, so outsourcing to CROs leaves the model "missing a leg." He says: "the experimental results are reward signals, used for reinforcement-learning post-training... that's why you must build your own lab" [01:12:13]. The economics echo GPU rent-vs-buy: "the cost of buying cards is roughly three times the cost of renting... run it more than three years and buying is clearly cheaper" [01:05:22]. He also adds that outsourced labs may not move at the speed scaling requires: "otherwise either you get choked or others may not cooperate with experiments that fast" [01:05:52].

A triggering event: Chai/Trillion-style antibody design (Try2) showed the technology had crossed a threshold

Gu says one result flipped his team's strategy. He refers to Try2 from "Try Discovery Lab" (garbled transcript name; likely Chai Discovery's Chai-2): "when we saw that model we decided to switch technical routes and go all-in" [00:14:06]. The signal was a 10%-20% antibody/nanobody design hit rate versus "the one-in-a-thousand success rate of David Baker's lab" [00:05:47]. His reading: "this model, built on something weaker than AlphaFold3, already designs well" [00:18:56].

Fast strategic pivot inside a big company: six months from signal to beating the benchmark

After Try2 appeared, the team pivoted immediately: "if we didn't make fast adjustments, it would mean our sense of smell wasn't sharp enough" [00:08:11]. "We started SeedFold/SeedPro2 in June last year, and by end of December we'd done it and surpassed them" [00:08:11]. SeedPro2 hit "70% to 80%" design success with binding affinity "several tens of times" Try2's [00:07:14].

Pretraining is an organizational capability, not a secret

Asked what separates top teams, Gu says it's execution and process: "your technical plan must be complete, no obvious mistakes... details must be done carefully" [00:35:30]. Cross-team checking is the mechanism: "cross-team collaboration is effectively a cross-validation, which is good" [00:35:00]. Pretraining at ByteDance Seed is split across model architecture, data, infra, and optimization/scaling, each with an accountable owner reporting to Wu Yonghui: "Somebody has to sign off... otherwise everyone shifts blame" [00:34:02]. Big-company advantage: "every time we run the formal task we've done very sufficient research... small teams don't have resources to research enough, they can only 'you only live once' and run it" [00:41:46].

Scale-up failures and the economics of training-run risk

Gu describes "training collapse" broadly (loss spikes, gradient blow-ups, or downstream metrics stagnating) and says many Chinese teams hit it in 2H 2025 when scaling up: "as far as I know, in the second half of last year, some domestic teams, big companies and startups, all more or less ran into this" [00:28:11]. A failed run on a few-hundred-billion-parameter model "may be tens of millions of dollars" [00:30:10]. The reassurance is scaling-law validation: "once you've got scaling laws figured out, 1T may be a threshold... training 2T or 5T or 10T should be no big problem" [00:30:38].

The AI-for-coding blind spot and the six-month lag

Gu explains why Chinese labs missed coding early: broad bets versus Anthropic's focus. "Anthropic did code from the start, but OpenAI simultaneously did Sora, ChatGPT, and many other things" [00:52:26]. Realization came in 2H 2025 and shipping took roughly six months: "six months is a relatively normal cycle" [00:54:19]. Same pattern for drug design: his own awareness came June 2025 and results followed by December.

Universities vs industry in the scaling era

Gu says large models "punctured" academia on compute, student mindset, and pedagogy. He says: "you only have to spend half an hour asking a model and you finally get it... this is something I want to learn and think is valuable" [01:17:33]. He still sees PhDs as valuable for "research taste" but would not push most people to do one. He also contrasts cultures: industry encourages head-to-head competition; academia is "一团和气" and the leaderboards ("Arena.ai") have turned it into "MMA matches" [01:24:50].

2. Contrarian Perspectives

Drug discovery should be treated like coding, not disease-by-disease

When asked which diseases will be solved first, Gu rejects the premise: "asking this shows your thinking is still stuck in traditional AI drug discovery... Claude Code isn't aimed at some language or one class" [01:10:15]. His model: "as long as there is a potential biological mechanism there, a strong model goes from target discovery to drug design to validation... a self-loop, self-closed loop" [01:10:44]. The implication is a horizontal, general-purpose drug-design model sold as an API, not a disease-focused pipeline.

Don't distill: shortcuts cap you at second place

Gu confirms ByteDance forbids training on distilled data (as of his departure) and defends it: "if you want to beat the strongest one, relying on distillation, you can only ever trail behind" [00:38:57]. His football analogy: "bringing in a bunch of foreign players may lift results short term, but it doesn't solve team-building" [00:37:29]. The real prize is figuring out how top labs generate high-quality data: "you'd have a method to continuously generate high-quality data; this is the most essential thing" [00:38:29].

Foundation models will not absorb vertical bio models

Gu argues multimodal frontier models won't swallow structure design: "can SeeDance be folded into Doubao? I think no" [01:01:30]. By analogy, a general model can't match Seedance 2 in video: "if you want effect better than Seedance2 in video generation, that's impossible" [01:02:29]. Protein design is "orthogonal" to the LLM base model because the underlying principle (3D structure) differs from sequence. This is a bet that vertical foundation-model companies survive.

Small molecules are a small market; large molecules are the trillion-dollar prize

Gu says: "I think the small-molecule space is already very small; large molecule is at least a trillion, possibly ten trillion market" [00:18:28], while acknowledging this contradicts intuition from high-volume cheap pills. His argument is about value: biologics treat cancer, leukemia, diabetes. He also stresses that a high-affinity small-molecule hit "is nowhere near a real drug" because of metabolism and multi-objective optimization [00:17:01]. Biologics break down into amino acids, so "you don't have to worry too much about metabolism" [00:17:28].

Competition from OpenAI/Anthropic entering your space is validation, not threat

Gu says: "I have no pressure; instead I'm excited because what I love most is competition" [00:13:07]. On John Jumper joining Anthropic: "their starting to lay out is exactly a validation or endorsement" [00:13:36]. And a rule of thumb: "if someone tells you they chose a track with only them and no one else, you need to have a second thought" [01:28:41].

"Pretraining is a bit boring now" and the sharpest people may need new challenges

Gu admits he found late-stage pretraining a little dull: "challenges got fewer" because options were already clear and team consensus took time [00:48:08]. He also pushes back on Yao Shunyu's "reliable is enough": "he's already very smart, so he's humble" [00:45:40]. His view is that originality is "innate": "this person has to be sharp enough, smart enough" [00:45:10].

3. Companies Identified

Geodesic Intelligence (Gu's new company)

A newly launched AI drug-discovery company (referred to in the intro) that aims to scale structure-based protein design on a foundation-model basis with an in-house wet lab. Why mentioned: Gu's startup, with an Anthropic-style API model. Quote: "our model is very similar to Anthropic, give pharma an API... they use my API and may also make good drugs" [01:07:47]. He also targets 2026 results: "if there's my company, [Kevin Weil's] prediction will certainly be fulfilled" [01:13:41].

Try Discovery Lab / Chai Discovery (the "Try2" model)

A startup (founders ex-Meta/OpenAI per Gu; the transcript garbles the name, likely Chai Discovery) whose Chai-2-like model designs antibodies and nanobodies with 10-20% success. Why mentioned: the catalyst for Gu's pivot. Quote: "our first reaction to Try2 was to switch to similar technical lines... and go all-in" [00:08:11]; "the founder is in AppSite... maybe Meta and OpenAI; it might have some roots with OpenAI" [00:14:36].

ByteDance Seed (and Seed team: SeedFold, SeedPro2, Seedance, Doubao)

ByteDance's foundation-model organization. Why mentioned: Gu's former employer; talent density. Quotes: "Seed's AI-native talent density is very high, extremely high" [00:09:39]; "Seed 2.1 we released Feb 14, that time it was world number one" [00:40:19]; coding rank claims "just behind Anthropic's best" [00:54:19].

SeedFold / SeedPro2 (Gu's team products)

SeedFold is a 900M-parameter structure-prediction model that beat AlphaFold3 on FoldingBench; SeedPro2 is a structure-design model. Quote: "SeedFold is the only one besides their own that dares to claim it beats AF3" [00:20:55]; "SeedPro2's design success rate reached 70% to 80%" [00:07:14].

Google DeepMind / AlphaFold

Maker of AlphaFold series. Why mentioned: baseline and surprise that DeepMind pursued small molecules rather than protein design. Quote: "DeepMind went off to do small-molecule design... it didn't do this large-molecule, protein design. This is what surprised me" [00:16:03].

Anthropic

Frontier lab; has hired John Jumper. Why mentioned: coding success and AI-for-science entry. Quote: "Anthropic's success in coding agent is so big, everyone must make some retrospective" [00:52:54].

OpenAI

Frontier lab with an AI-for-Science push (Kevin Weil). Why mentioned: validates AI for science; broad bets. Quote: "AI for Science 2026 will be to 2025's AI for software programming" (Kevin Weil's prediction, restated by host) [01:13:41].

Bose (Boltz) / BoseJump / Bandercraft (Baker's student team) / Latent Labs

Other teams building structure-design models. Why mentioned: peers in Try2-style models. Quote: "BoseJump, a team of MIT background; David Baker's student team, Bandercraft... these all are the same class as Try2" [00:19:25]. Gu on peers: "real ones who can train their own models are actually very few" [01:09:46], and "among them only one or two will make it out, like the company in the past" [01:10:15].

Meta (ESM)

Protein language model benchmark. Quote: "our model is also better than ESM" [00:59:04].

DeepSeek

Why mentioned: first with open-source V3/R1 (670B+) that reset ByteDance's scaling goals, and Gu's critique of its architecture. Quotes: "DeepSeek V3, R1 at that time already trained 670B models and ours were far from that scale" [00:25:16]; "mHC, Spot Attention DSA, to me not elegant enough, structure too complex... not good for scaling" [01:19:00].

Cursor / Claude Code

Why mentioned: coding agents as the template for what bio could become. Quote: "in the next one or two, two or three years there will be a product like coding agent or Claude Code" [00:14:06].

NVIDIA (Cosmos 3)

Why mentioned: example of unified multi-modal model for robotics/physical world. Quote: "recently NVIDIA released a model called cosmos3... focused on robotics and physical-world generation" [01:02:00].

Chinese AI-bio incumbents: XtalPi (晶泰), Insilico (英矽智能)

Why mentioned: traditional biotech with running pipelines. Quote: "in China there's Yingxi Intelligence, Jingtai, their experiments and pipelines run well" [00:56:41].

Asimov / Lila-type CRO labs

Gu references labs that rent experimental capacity ("CIO"/CRO-like) as comparisons to selling compute. Quote: "they become a 'lab cloud', like serving models" [01:04:25].

Ex-ByteDance spinout "A-LiuLibs"

Quote: "Liu Kai set up a company separately... still a main shareholder" [00:04:18].

4. People Identified

Gu Quanquan (顾全全)

Founder of Geodesic Intelligence; ex-ByteDance Seed (AI drug team lead, pretraining optimization/scaling co-lead), ex-professor (machine-learning theory, UCLA-based context). Why mentioned: guest. Quote: "I spent three years turning myself into an industry person" [01:06:49].

Wu Yonghui (吴永辉)

ByteDance Seed head, joined 2025. Why mentioned: orchestrated cross-team pretraining. Quote: "Yonghui very much encourages cross-team collaboration, and since he came the atmosphere and efficiency have improved a lot" [00:33:02]; "he also grasps the details well" [00:35:30].

Xiao Wenzhi (肖文芝 / "Wenzhi")

Led ByteDance's Protenix (AF3 reproduction); ex-search/ads/recommendations; left earlier. Quote: "Protenix model, in the field everyone recognizes it" [00:05:17].

Liu Kai (刘凯)

Ex-IDG/Volcano investor in healthcare; led ByteDance small-molecule drug team, now founder of A-LiuLibs. Quote: "he used to be an investor at IDG and Volcano Engine's... medical-focused investment" [00:10:09].

John Jumper

AlphaFold series author who joined Anthropic. Quote: "John Jumper joined Anthropic... he's the author of the AlphaFold series" (host) [00:12:37]; Gu says this makes him "excited, no pressure" [00:13:07].

David Baker

Protein-design lab head. Quote: "success rate of one in a thousand of David Baker's lab" [00:05:47].

Liang Rubo (梁汝波)

ByteDance CEO; Feb 2025 all-hands set three goals (intelligence ceiling, new interaction, scale effect) and reflected on being slow to follow OpenAI's route. Quote (host paraphrasing): "he reflected... following the O1 route was not fast enough" [00:51:28].

Yao Shunyu

Researcher cited for saying one doesn't need to be very smart, being reliable suffices. Gu's response: "he's very smart and sharp himself, so he's humble" [00:45:40].

Kevin Weil

Head of OpenAI for Science. Quote (host): "he said 2026's AI for Science will be to 2025's AI for software programming" [01:13:41].

Ilya Sutskever

Quoted by the host on "air being sucked out by LLMs and scaling laws"; Gu's response is that scaling is general [01:29:10].

Try2 founders (ex-Meta/OpenAI per Gu)

Quote: "the founder previously was at... maybe Meta and OpenAI both" [00:14:36].

5. Operating Insights

Institute accountable owners with cross-team "closed-loop" challenge

Gu explains why pretraining has separate teams with named owners: "each person must sign... this data is mine, so I must ensure it's better than the last version" [00:34:02]. Combined with open challenge: "if just one team, only they say it, no one challenges you, easier to make mistakes" [00:35:00]. Operators can replicate: split a complex pipeline into owner-signed modules plus a ritual of cross-challenge, with a single arbiter for disagreements.

Build consensus even if you're in the 10% who already see the answer

Gu says waiting is the correct but boring discipline: "only when 90% of people get on the same page can you move to the next step... you can't just rush ahead and let the team not fully understand" [00:49:05]. For leaders, the lesson is that premature unilateral action costs team buy-in and contribution credit: "everyone wants their own contribution" [00:49:32].

Don't flip-flop between technical routes; hold your own coherent worldview

Gu warns against "人云亦云": "you can't hop between schemes... you might end up with nothing or a hybrid mess" [00:43:38]. Hiring heuristic too: "people who've seen many pigs run don't necessarily eat pork" [00:43:11]; hiring on "worked at famous company X" is "utilitarian and short-term" [00:45:10].

Use small-model research to extrapolate: verify scaling law at 1T before going bigger

He describes the transfer discipline: run research on small models, fit scaling laws, validate at a large point, then extrapolate "20-50×" [00:31:34]. An operational rule: once the predicted loss matches actual at the largest tested size, risk of further extrapolation failure is low [00:31:07].

Pivot fast when a competitor's result invalidates your path, even at the cost of overlap

Seeing Try2, Gu's team dropped its plan and went to a Try2-style route in six months, accepting overlap with another internal team: "that's unavoidable... if we got beaten by Try2, that's not good for the company either" [00:08:40].

6. Overlooked Insights

Pretraining and post-training are decoupled in a way that hides failure attribution

Gu says in passing that when post-training stalls, nobody can tell whether pretraining was bad, was too good to improve further, or was misaligned with post-training: "to figure it out you'd have to put all the pretraining and post-training people together" [00:50:29]. The significance: the next structural improvement may be tight pretrain/post-train co-design, and the org seam is a hidden source of lost model quality. Whoever integrates these two stages (organizationally and technically) could gain an edge.

The wet-lab reward loop turns biology into an RL environment, which pretrained-only competitors can't copy

Almost as an aside Gu says pretraining needs essentially no experiments but "post-training does, experimental results returned as reward signals" for affinity, hit rate, toxicity, synthesizability, specificity [01:12:13]. This implies a proprietary data flywheel: each experiment improves the model, and a company with its own lab compounds while a pure-compute or pure-CRO rival does not. It also explains why his compute-heavy pitch includes a mere "few million dollars" lab line item, which makes the moat cheap relative to its value.