Two honesty notes up front.
- Adaption Labs (adaptionlabs.ai): I have low-confidence info on the founders and funding. Public details are thin — slick site, "Building Intelligence that Evolves with the World" tagline. Treat everything below about them as unverified; use the LinkedIn/Crunchbase checks listed at the end before acting.
- Jerry Tworek: he was VP of Research at OpenAI (reasoning / agents / o-series lineage). I do not have a confirmed source that he has left to start a new company. If you've seen credible reporting (a tweet, a TechCrunch piece), tell me the source so I can include it accurately. I'd rather flag the gap than fabricate a founding story.
1. Why continual learning is the right bet
The premise: today's frontier models are frozen after pretraining + post-training. They learn nothing from interaction. Every "in-context" learning trick is a workaround for the fact that weights don't update. The bottleneck arguments:
- Sutton's bitter lesson update — his Alberta Plan and recent essays argue intelligence requires continual learning from the stream of experience, not curated batch training. Frontier labs are largely ignoring this.
- Catastrophic forgetting — when you fine-tune a model on new data, it loses prior capabilities. Solving this cleanly is unsolved at scale.
- Sample efficiency — humans learn from one example; LLMs need millions. Closing this gap is closer to "intelligence" than scaling parameters.
- Why it matters for investing — markets are non-stationary. A model frozen in 2024 cannot incorporate that the regime shifted in 2026. An investor-AI must be a continual-learning system, or you're just doing fine-tunes forever.
The contrarian counter: most frontier-lab researchers think continual learning is solvable inside the current paradigm via retrieval + episodic memory + periodic re-training, not as a separate research agenda. If they're right, dedicated CL startups are betting on a problem that big labs will absorb. If Sutton is right, CL startups have the moat. This is a real research disagreement, not a settled question.
2. The companies (what's actually there)
| Company | Approach (claimed) | Founders | Stage / Val | P(meaningful 5y) | Notes |
|---|---|---|---|---|---|
| Adaption Labs adaptionlabs.ai |
"Intelligence that evolves with the world" — continual / online learning. Specifics not public from what I can find. | Unverified. Founders not confirmed in my sources. | Likely seed/A; unverified | 15% | Strongest aesthetic on the website. Lots of "fundamental innovation" marketing, light on technical claims publicly. Could be a serious deep-tech bet or a vibe play — need primary sources to tell. Probability is "wide error bar" — could be 5% or 35%. |
| Jerry Tworek's startup if it exists |
Unknown. If real, almost certainly RL + reasoning continuation given his OpenAI background. | Jerry Tworek (ex OpenAI VP Research, reasoning team) | Unverified | ? | If verified, this would be a top-tier bet — Tworek is one of the most senior reasoning researchers in the world. Send me the source and I'll re-rate properly. |
| Sakana AI | Evolutionary methods for model improvement, agent collectives, self-improving systems. Not strictly CL, but adjacent (model-evolves-itself). | David Ha, Llion Jones (ex-Google, Transformer co-author) | ~$1.5B, Tokyo | 35% | Real research, real publications. Tokyo HQ — bad for Seattle bet, possibly remote-friendly for senior researchers. Closest thing to a "continual self-improvement" lab with a real track record. |
| Liquid AI | Liquid neural networks (continuous-time dynamics, fewer params). MIT spinout. Architecturally adjacent to CL — built for streaming inputs. | Ramin Hasani, Mathias Lechner, Daniela Rus (MIT CSAIL) | ~$2B, Boston | 25% | Real architecture work, real defense-industry interest. Less directly "CL" than they market themselves. Boston, not Seattle. |
| Manifest AI / xLSTM offshoots | Architectures that handle long horizons + state better than transformers (xLSTM, Mamba derivatives, RWKV). State-space models naturally handle continual streams. | Sepp Hochreiter (xLSTM), Albert Gu (Mamba, Cartesia), various | Cartesia ~$1B; others smaller | 30% | Cartesia is the most fundable — Albert Gu + voice product. Architecture itself doesn't equal CL but is a substrate. |
| Numenta | Jeff Hawkins' brain-inspired CL (HTM, Thousand Brains). Long-running research org now selling sparse-AI inference acceleration. | Jeff Hawkins, Subutai Ahmad | Private, mature; pivoted to inference | 10% | 20+ years on this thesis without a commercial breakthrough. Cautionary tale — being early on the right idea is not enough. |
| Verses AI | Active inference / Karl Friston's free energy principle. Bayesian agents that update online. | Karl Friston (chief scientist) | Public (small cap) | 5% | Strong scientific advisor, weak commercial execution so far. Public stock = no startup-equity upside. |
| Periodic Labs | RL on real-world experiments (science). Continual learning from physical trial-and-error. | Liam Fedus (ex OpenAI), Ekin Dogus Cubuk (ex Google) | Seed/A, SF | 35% | Already on your main list. Methodologically the closest to "AI scientist that updates from new evidence" — the same shape as your investing thesis. |
| Cohere / Mistral / smaller frontier labs | Periodic re-training pipelines, retrieval-augmented systems. Not CL in the strict sense. | — | Various | — | Listed for completeness — they are not pursuing the CL research agenda as a thesis. |
| Google DeepMind (CL group) | Real continual / lifelong learning research lineage (EWC, PNN, etc.). Multiple senior researchers. | — | Public (GOOG) | 90% | If you want to do real CL research at scale, DeepMind has the deepest bench. RSU not startup equity. No Seattle. |
3. The honest probability picture
Continual-learning startups are even harder bets than chip startups, for two reasons:
- The science isn't settled. No one has demonstrated a CL system that scales to GPT-4-class capability without catastrophic forgetting. You could join Adaption Labs and find out in 18 months they hit the same wall everyone else has.
- Frontier labs may eat the niche. If continual learning turns out to be solvable with "just keep training on new data + replay buffer," then OpenAI/Anthropic/DeepMind win by default and CL startups become acquihires.
That said, the upside is uniquely large: if a CL startup actually cracks lifelong learning at frontier scale, that is the AGI breakthrough. So the bet is binary in a way that's almost ideal for "$1M equity → 100x" math — but the base rate is closer to 5-15% per company, not 30%.
4. To work at one of these — what it takes
Skills they actually want
- RL fundamentals: PPO, GRPO, off-policy methods, exploration. Sutton & Barto cover-to-cover, Spinning Up, recent RLHF / RLAIF work.
- Memory architectures: external memory networks, retrieval, episodic memory systems, Atlas / TITAN / RecurrentGPT-style work.
- Catastrophic forgetting literature: EWC (Kirkpatrick et al), Progressive Networks, Synaptic Intelligence, replay-based methods, more recent LoRA-based CL.
- State-space models: Mamba, Mamba-2, xLSTM, RWKV — architecturally relevant.
- Empirical depth: at least one project where you ran a non-trivial training experiment and reported on a distribution shift / forgetting metric. This is the single thing that distinguishes "interested" from "credible."
Concrete prep plan (3-6 months)
- Reproduce one canonical CL benchmark — Split-CIFAR, Permuted MNIST, or better, a streaming-LM eval. Compare 2-3 methods (naive fine-tune, EWC, replay). Write it up. This is your résumé piece.
- Write a public position essay: "Why continual learning is the missing piece for AI investors." Connect Sutton's argument to your investing thesis. Founders read essays; this is your introduction.
- Read & annotate: Sutton's Alberta Plan, "A Definition of Continual Reinforcement Learning" (Abel et al), the Mamba papers, Anthropic's "Towards Monosemanticity" (relevant for reasoning about what a model knows), and any recent paper from Sakana AI.
- Map the people — for each of the top 3 CL companies (Adaption, Sakana, Periodic), identify the 5 most senior researchers on LinkedIn. Reach out with the artifact + essay.
- Talk to Sutton-adjacent academics — Marlos C. Machado, Doina Precup, Marc Bellemare. Real CL research is concentrated in a small Alberta/DeepMind/MIT network. Knowing the people is half the entry.
How offers work here
CL startups are smaller and more research-driven than chip startups. Realistic comp at seed/A: $300-450K base + 0.1-0.5% equity for a senior IC. Founding-engineer-level roles (much rarer, much harder bar) can be 1%+. The $1M-equity-that-100x's math works at this stage if the company hits — at $30M post, 0.3% = $90K nominal, which goes to $9M at $3B. That's the realistic shape of the bet.
5. My honest take
Best fit for you, in order:
- Periodic Labs — already on your main list. Closest to your investing thesis (AI updates beliefs from real-world evidence). Strong founders, ex-OpenAI/Google. Seed/A means real equity %. SF — not Seattle.
- Sakana AI — strongest research credibility in the "model self-improves" space. Tokyo HQ is a real obstacle; check if they have remote senior roles or a US satellite.
- Adaption Labs — verify before pursuing. If founders are credible (check LinkedIn, Crunchbase, recent funding announcements), this could be a top pick because it's earlier than Periodic. If founders are weaker, skip.
- Jerry Tworek's company — if it's real, almost auto-include at the top. Send me a source.
- DeepMind continual learning team — if you want to do real CL research with no equity upside, this is the place. Bad for the $1M-100x bet, good for credibility-building.
The contrarian play: instead of joining a CL startup, work at a frontier lab on memory / retrieval / RL infrastructure for 2-3 years, then start your own CL company aimed specifically at investing. Continual learning + your investing domain expertise is a very narrow founder profile and a real wedge. The "BUILD YOUR OWN" row on the main table isn't a joke for this category.
6. Verification checklist for Adaption Labs
- LinkedIn search "Adaption Labs" — who are the founders, who has joined, what's their pedigree
- Crunchbase — funding round, lead investor (a16z / Khosla / DST signals real, smaller signals less)
- Their Research and Blog pages on adaptionlabs.ai — are there technical posts or just marketing?
- arXiv search for the founders' names — do they have CL publications or is this their first AI rodeo?
- Glassdoor / Levels.fyi — any comp data? (Likely none yet if early.)
Send me what you find on items 1-2 and I'll write a real entry for them with a calibrated probability instead of a placeholder.