> I think agent insurance is coming. Once an agent makes an expensive mistake, insurers will want audits before they cover anyone.
I spent 48 hours in SF hanging out with billionaires, data labs and AI founders and execs running $1B to $100B companies.
My observations:
1. Sooo MANY stories of data brokers making like $2M a WEEK selling company data to labs. I can't tell if it's the beginning or a bubble, and I kept wondering which niches are left. Digging into it actively.
2. At dinner, someone asked the table, "If a new model drops next month, does your company get better or worse?" HALF the table went quiet! I think that's the best question to ask about any AI startup right now.
3. The bar people talk about now is hitting $1B in revenue in 12 months (I guess that's the new unicorn?). What I didn't hear much about is retention. AI revenue can come in fast and leave fast, and I think we're going to see a lot of stories of AI companies with huge revenue and terrible retention.
4. It's pretty clear we're living through the reinvention of the computer. The model is the CPU, the harness is the operating system, and skills and plugins are the apps. Thinking about it that way makes it easier to see where the whitespace is.
5. We're maybe 12 months from an agent that can watch you work for a week and then do your job without any instructions. The screen recording plus agent memory plus local model combination makes this possible right now.
6. Agent harnesses feel like the new GPT wrappers, in a good way, especially vertical ones. The best teams I talked to treat the model as swappable. One founder walked me through routing different models to different jobs: one for taste and visuals, a cheaper one for the everyday stuff, another for text.
They use prompts they can move across models over fine-tuning, and A/B test the output on what users actually do with it: do they edit it, export it, share it? A vertical harness is basically four things: the connectors for that industry's data, evals written by people who know the work, rules for what the agent can and can't touch, and a clean screen for a human to review. The model is maybe 20% of it.
7. There's a lot of reverse engineering going on/cloning going on more than I've ever seen. People are decompiling software and rebuilding it with agents, and someone showed a suite of like 10 apps that are worth $5B+ each all that he cloned open source. Me and everyone are still thinking through what it means when software gets this easy to copy.
8. My guess is the value moves to data. If the software can be rebuilt quickly, what's worth paying for in a legacy SaaS company is the years of customer data and workflows inside it. Some SaaS companies are secretly data companies.
Gets me thinking about how do you truly value data for this saas companies? Like if we were buying these dying saas companies at 2x ARR before (as example), maybe there is a better metric. Idk, just thinking out loud here.
9. It feels like a good chunk of legacy software won't make it to the agentic era, especially anything built around people clicking through a lot of screens.
10. Every company I talked to wants to pipe its internal data into one place and expose it to agents through MCP. Then they look at the data and realize it's a mess: five versions of the same customer, old Slack channels, PDFs nobody can search. Cleaning that up feels like a big consulting opportunity!
11. A lot of founders are building the basic building blocks for agents: APIs, data layers and tools. Some of the companies I saw barely have a UI because the main user is an agent. I really believe the startups that build primitives to this agentic era will do extremely well. Building blocks is the best way to think about it.
12. Labs can get plenty of training data, but it's harder to get real examples of agents failing inside companies. Whoever has those logs has something valuable.
13. Someone running a big consumer company told me the next fight is over memory. The model that remembers the most about you wins, and the smartest model matters less than people think.
14. Physical AI came up a lot. Astra I think was the first model that got people into AI can actually do CAD, and I think a lot of people are thinking what does vibe manufacturing really look like. The line I keep thinking about came from a robotics founder: the hardest problems in AI are deployment problems now, not research problems. Their robots work on their own, call a human when they get stuck, and every one of those calls becomes training data.
15. Cybersecurity companies are having a moment and are actually cool companies to build now, especially the ones that clean up after a breach. Every agent with access to your systems is a new door into your company, and a lot of those doors were left open.
16. Some startups are worth more dead than alive now. A shutdown company with 5 years of Slack, email and code can get paid for licensing all of it. I wouldn't be surprised if founders start factoring that into when they shut down.
17. People are starting to talk about monetizing the agent web. It feels really early, but when the most popular user of the internet is an agent, what startups should you be building?
18. I think agent insurance is coming. Once an agent makes an expensive mistake, insurers will want audits before they cover anyone.
19. Several people said the hire they want most is someone with taste who can write evals. The models are getting good at generating ideas, so the hard part is judging which ones are good.
20. New trend I observed: startup raises VC, uses venture dollars to become a creator, startup fails, founder falls back on being a full-time creator and incubates other products bootstrapped.
21. Getting SIGNIFICANT distribution is HARD right now. Founders are sponsoring buses, buying billboards like crazy and throwing parties with famous DJs to get noticed. Kinda feels dotcom era in that sense.
Attention is what's scarce! The growth guy going viral sharing his UGC strategy from Wispr Flow is a good example. It's the sauce everyone wants, and I truly believe there are probably only 50 people on the planet who know how to scale these types of businesses efficiently in this era.
Always good to get that dose of SF. I felt sooo lucky to have such an incredible trip during such an incredible time in building businesses. Surreal in so many ways.
Happy to be home and locked in.
Until next time!
Nikola G retweeted
yup, curriculum learning is probably as old as RL :D
People usually think of RL as "the agent collects its own data", but the robot can't really pick its own data to solve a hard exploration problem, and see useful gradient signal, if the chance of seeing success is super low (in a sparse reward setting). Which then leads to using denser rewards so that the gradient signal is not impossibly rare to stumble upon. Which then leads to tons of reward tuning and task-specific curriculums. Which then leads to suboptimal policies which are optimizing wrt a soup of mishmashed objectives. Which turns out leads to an inability to scale. So the details of SGS that matter are:
1. Expose a ton of "task configurations" (initial state, goal) to the robot, so that the difficulty of the set of tasks the robot can actually pick from is actually diverse.
2. Instantiate a very simple, automatic, adaptive curriculum based on success rate of each of these task configs (SGS) so that the robot doesn't waste experience in tasks that are neither too easy nor too hard.
So the surprising thing is that this simple instanciation lets us train policies with a sparse reward formulation, at much higher scale than previously possible, to achieve really hard robotic tasks :D
Nikola G retweeted
If you've ever worked at Google you'd see first hand just how much raw cognitive math-ish processing speed and working memory can decline for nationally ranked test takers and former top uni grads in their 30s and 40s, from decades of disuse and unambitious office work.
Nikola G retweeted
What GPT6-Astra tells us is that scaling spatial and 3D understanding rather than scaling action leads to better robotic manipulation.
We take a minimalist approach: what if we train a policy directly from a 3D understading model, not a VLM? Surprisingly, it outperforms most larger VLAs and WAMs.
Introducing Grounding Action Model—a brand new and SOTA policy train from a small 3D box detector WildDet3D（allenai.github.io/WildDet3D/）
Arxiv: arxiv.org/abs/2609.23863
Code: grounded-action-model.github…
Nikola G retweeted
After 2+ years in the robotics data space, we are shutting @Eidon_AI down.
The thesis was right. But the business is brutally hard.
We close this chapter by open-sourcing everything we built and sharing lessons for anyone venturing into the space.
Nikola G retweeted
A carefully controlled look at looped transformers (arXiv 2609.19107):
1. Weight sharing is not compute-optimal on fresh data. It costs a constant ~1.06–1.16x compute at every scale.
2. The looped model's gains over vanilla come from elsewhere: the norm + input re-injection boundary operator, and growing depth mid-training. Both improve the scaling exponent. Neither needs weight sharing.
3. Looping is still a parameter-free way to get growth. Tied growth keeps the exponent gain at vanilla param count.
4. Weight sharing wins on multi-epoch, data-constrained training.
arxiv.org/abs/2609.19107
What if the underlying tokens of images are actually particles represented as points?
Thanks to awesome co-authors @kaiwynd @Adamjhung @bowenwen_me @BirchfieldStan @YunzhuLiYZ @RamananDeva @jeff_ichnowski
The paper is out now; the code and dataset will follow shortly.
Website: pointzero-wm.github.io
arXiv: arxiv.org/abs/2609.19142
6/6
Nikola G retweeted
Maybe psychosis kicking in here, but this is absolutely absurd the more I think about it. OpenAI compressed 100 years of mathmatical research into 88 hours and it cost them let’s say 10 million dollars.
So theoretically if you have a problem thag can be broken down into man hours, we essentially have an estimated cost to solve them.
That’s absurd. How many problems have a price tag hanging on them after this?
Proof by counterexample strikes again
We’re sharing a solution to the Navier-Stokes Millennium Prize Problem, one of the deepest problems at the frontier of mathematics.
The proof was produced by a group of agents, using an OpenAI next-generation model significantly more capable than GPT-6 Astra.
The problem concerns whether the description of smooth three-dimensional fluid motion modeled by the Navier-Stokes equations can break down. It has remained unresolved for roughly 90 years.
Nikola G retweeted
Commenters on HN are uncovering more wikis and public sites apparently used by OpenAI agents to communicate on the open web.
Despite read-only web access, the agents were able to leave ~18,000 posts sharing answers and bypasses. But now, users are discovering more.
This appears to be a separate swarm from the one that attacked Hugging Face.
news.ycombinator.com/item?id…
Nikola G retweeted
Seeing a hype wave around GEN-1.5, and rightfully so. Lots of respect to Pete & Andy for executing so well. The secret is in the naturally repetitive motions in human-collected data. There're 2 main sources for such repetitions:
(1) Symmetric patterns. Sorting, tidying, and assembling almost never finish in one motion. Open any assembly manual from IKEA, and you find most objects symmetrical. You drive one bolt, then its twin, then the next pair. Every {bolt A, bolt B} pair is a natural continuation in context, and the second instance is a free training signal that imitates the first ("prompt").
(2) Recovery. Humans drop things all the time, but we pick them up so fast, we don’t even notice. That reflex to fix is half of our physical competence. The key insight is to keep the failed first half instead of trimming it away. If the model consumes the full arc, fumble, catch, continue, then recovery shows up organically at test time. It's funny that in-context improvement results from *NOT* over-sanitizing your data.
The other critical ingredient is UMI. I've been saying for a while that teleop will not last, and GEN-1.5 is driving the final nail in the coffin. UMI is essentially a human wearing the robot gripper to collect data directly (human → data). Teleop inserts a layer of separation: human → VR/skeletal device → robot → data, which bleeds out all the human "physical intuition". The subtle sleight of hand we perform constantly with objects, the micro-adjustments, the feel of a part snapping into place, is nearly impossible to capture when you can't feel the environment directly.
Once you have enough data, many behaviors can actually be zero-shot. For example, you don't even need finetuning to pick up a novel object. The model "just knows" what to do given a similar scene in the training distribution. Whether in-context learning truly works or not also depends on how far away the test is from training. Currently, the demos are still a bit too simple to conclude.
I'm cautiously optimistic. Still, it's a great day in robotics.
Reading about ancient history, it struck me that successful civilizations avoid over-optimizing for the present. Laws that protect the weak from the strong preserve people, skills and ways of life that may seem marginal today but become essential when the environment changes.
> only larger models, by virtue of having more parameters and hence less gradient interference, are able to retain memory of a previously observed batch of data from a rare task... In contrast, the smaller model is forced to start from scratch and consequently fails
We take for granted that larger models are better than smaller ones, but why is this so? Our new paper, led by Jing Huang and @EkdeepL, traces this to a data-induced competition for resources (neurons), using formal analysis, idealized tasks, and real pretraining.
ALT Title card for a research paper. The title reads "Why Larger Models Learn More: Effects of Capacity, Interference, and Rare-Task Retention." Authors listed: Jing Huang, Daniel Wurgaft, Rachit Bansal, Laura Ruis, Naomi Saphra, David Alvarez-Melis, Andrew Lampinen, Christopher Potts, and Ekdeep Singh Lubana. A Goodfire logo appears below the names. Author affiliations: Stanford University, Kempner Institute at Harvard University, MIT, and Anthropic.
Dug deeper into this and realized the algorithm they use is essentially online PCA. There is a nice survey on it here: arxiv.org/pdf/1511.03688
Starting to think "loss of plasticity" is an intuition that rarely occurs in practice.
Haven't managed to get a model to lose plasticity at all. Instead they suffer from catastrophic forgetting much more rapidly.
