Fellows Forum

A private annual gathering of the founders, researchers, and investors shaping the future of AI and robotics, hosted by @Fellows_Fund

[Fellows Forum @Fellows_Forum]

.@pirroh’s on Replit’s B2C to B2B transition:

“I believe most of the AI companies will be way more hybrid than the old school playbook where investors told you you're either a B2C or a B2B company. I think that paradigm is gone.”

[Fellows Forum @Fellows_Forum]

.@MillionInt led the reasoning research behind OpenAI’s o1 and o3 models before leaving to start @coreauto.

Here are some of the best things he said at Fellows Forum 2026:

1/ “OpenAI, Anthropic and a few others training LLMs already have a very successful recipe. In some ways it prints money right now. But changing a recipe that successful is hard, and I have a researcher’s spider sense that there are better ways to train large neural networks. That’s a risky bet that’s easier to take in a new company.”

2/ “We can max out basically every benchmark in the world, and models still fail a lot. Why is a model that’s solving Millennium Prize problems still failing at my day-to-day work?”

3/ “The Transformer is due for a refresh. We should be grateful for how far it’s gotten us, but if you ask me where the bottleneck is to our models getting much better, that’s where it lies.”

4/ “Today’s LLMs know everything about every footballer, every king and queen of England, every little village in Poland, but very little about you, me, my friends and what I did yesterday. What if we reversed it? A model that knows me, my work and my friends intimately is a very different model.”

5/ “Backpropagation has two fundamental problems nobody has solved. Gradient descent isn’t data efficient, and it causes catastrophic forgetting. If I hurt myself once, I don’t want to have to hurt myself a hundred times to learn from it.”

[Fellows Forum @Fellows_Forum]

.@ajay_bcv (BCV) on the robotics vertical he’s most bullish on:

“We’ve made a few investments in construction, and we think it’s a massive vertical. Data centers, infrastructure, residential, multifamily: it’s a huge opportunity to address labor shortages, the housing crisis and infrastructure demand. It’s one of the largest industries in the world and it’s still largely manual. Drive past any construction site and you’ll see equipment sitting unused 90% of the time. We think there are a dozen companies to be built here.”

[Fellows Forum @Fellows_Forum]

Mahi de Silva on one of Higgsfield’s early pivots:

“About a year into the company's inception, we killed the apps. We took them out of the app stores. We did a hard reset towards desktop and web apps, and that's when the business really took off.”

[Fellows Forum @Fellows_Forum]

TL;DR of our drug discovery panel w/ Hetu Kamichetty (Xaira), @exnx (Radical Numerics), and our AI Fellow Vijay Narayanan (SAIGroup):

-Why biology needs its own frontier models
-DNA, RNA and proteins as context
-The bottleneck after the bottleneck (and the FDA)
-Detecting deepfake viruses
-Where the moat moves: speed

Watch the full panel: https://t.co/cJ5FiBOg1Q

[Fellows Forum @Fellows_Forum]

Li Jing, ex-OpenAI and currently in stealth, when asked if world models need 3D data:

“I think for the world model itself, we definitely don’t need 3D. When we did Sora, at the beginning, everyone suspected we used so much 3D data. We actually used zero. It’s purely video generation. Eventually it learns all the 3D by itself.”

[Fellows Forum @Fellows_Forum]

.@MatthewPVail (OpenAI) asked our VC panel what they want founders to build next, given releases like Astra:

Our GP @AlexFellowsFund wants to see a robotics company building its model on GPT-6 capabilities.

@Jamesflyy (Sequoia) wants a consumer AI doctor. He sees ~95% of healthcare as an access problem that AI could solve cheaply, if a company gets regulatory approval and does it responsibly.

@Nina_Gerson (Index) sees an opportunity to rewrite the entire app stack. She thinks founders and investors fled the app layer too early.

@rodriscoll (Scale) dodged the question…we’re still waiting on a straight answer Rory.

[Fellows Forum @Fellows_Forum]

Hetu Kamichetty (Xaira) on where AI biosecurity regulation should sit:

“I am not of the opinion that the models can themselves be regulated. It’s just too easy for someone with a few GPUs to train a decent model. What is nice is that it took us months to make a vaccine for COVID when it came out. Today, it probably will take us weeks to days. So, just as bad things can happen faster, good things can happen faster too. The moment somebody wants to order DNA to make the molecule, you can introduce safety regulation.”

[Fellows Forum @Fellows_Forum]

Audience members were firing off great questions during this year’s Fellows Forum.

One member asked our physical AI panel if you need an external safety and coordination layer when multiple robots share a work cell. The speakers’ thoughts:

@nikhilabm (CloudChef) says most customers won’t let you deploy mobile robots without verifiable safety you can audit. He hopes a general-purpose company builds it. For now, they build it in-house.

@joshgruenstein (Tutor Intelligence) is skeptical of any third-party orchestration layer, since “there are 800 different parties in this ecosystem that all want to be it.” He expects it to sit with the customer or, more likely, the robotics company.

@adnothing (Walden) wants a Star Wars world, not an I, Robot one: many vendors, humans and machines sharing a floor, not a fleet of clones. Safety standards set the minimum bar, but each customer’s real bar depends on their risk tolerance and brand.

[Fellows Forum @Fellows_Forum]

At Fellows Forum 2026, @marcboroditsky (Nebius) told us something he’d never said before:

“Today it’s a bit of a best-of-breed reality in the market, where people are cobbling together a compute provider, a model-building capability, a post-training capability, an inferencing capability.

These are features of an intelligence platform. And my confidence is that when we talk about inference, it’s just a feature. It’s the API through which you get access to the model productivity, access to the intelligence. So my confidence is, based on success and momentum in the market, we’ll be selling an intelligence platform as opposed to just selling inferencing.”

[Fellows Forum @Fellows_Forum]

.@adnothing (Walden Robotics) on the standards robots are held to:

"There's a double standard between humans and machines. People tell you you can do it this way because humans do it this way. You deploy, and then they tell you the robot cannot do it this way, the robot needs to do it 10x better.”

[Fellows Forum @Fellows_Forum]

Mahi de Silva when asked if the studios will become a bigger part of the business:

"Every major studio uses our platform today. They're very shy about talking about that, and I'm not here to out them.

The reason is they're still having to struggle through the guilds and the unions, who I would say have a right to be concerned about what AI does to their craft.”

Give us some names at the next forum Mahi!

[Fellows Forum @Fellows_Forum]

Waymo's @AnguelovDrago asked our physical AI infra panel what role simulation should play in training robots, compared to real-world data:

Vasu Sundarababu (Centific) sees it as a force multiplier. You need some real-world evidence to start from, and it doesn’t replace real data.

@jingli9111 (ex-OpenAI) says training is about scaling up data & model to cover the long tail, and that human-designed sim scenarios never will.

@ericryanchan (Rhoda AI) thinks its first impact is in evaluation, and once you can evaluate, you can start training on it.

@olivercameron (Odyssey) believes it’s essential for evaluating safety-critical systems, and that world models are closing the gap between sim and the real world.

Dragomir himself says Waymo uses it to validate its systems, because you can't easily stage safety-critical driving situations in the real world.

Watch the whole panel here: https://t.co/kjQbVWULXY

[Fellows Forum @Fellows_Forum]

.@LiamFedus on their Neon model and the frontier XRD benchmark:

"We took a base model that started at basically a 3% success rate on our out of distribution tasks - effectively completely unusable. Through our mid-training and reinforcement learning, we were able to create a system that now exceeds frontier level.”

[Fellows Forum @Fellows_Forum]

7/10 largest food and beverage manufacturers in the world are running Laminar.

Here’s a quick recap of @annieJlu's keynote at Fellows Forum 2026:

-Process manufacturing, the other half of global manufacturing
-Chemistry intelligence
-Factories running on hard code
-Sensors that read chemistry inline
-Closed-loop control

Watch her whole speech here: https://t.co/qdLwogQvRd

[Fellows Forum @Fellows_Forum]

.@BorisMPower on the releases in between model vintages:

“A lot of those investments, the way that they're seen within the company, are extremely short-sighted. We do them because they help us iterate and learn faster today, not because that's the right long-term strategy. So 80 to 90% of the research at OpenAI is still focused at GPT-7, GPT-8 and beyond”

[Fellows Forum @Fellows_Forum]

.@peteflorence is one of the most impressive founders we’ve ever met.

Here’s the TL;DR of his keynote at Fellows Forum 2026:

-Scaling laws for robotics
-99%+ success from ~1 hour of task data
-One-shot learning from a single demo
-Zero-shot tool use
-Training across 9,000+ robot hands

Watch his full speech here: https://t.co/rO023CAIAt

[Fellows Forum @Fellows_Forum]

.@chuer_pan asked our robotics investing panel at Fellows Forum 2026 how much of the stack a robotics company should own:

Conner Lovely (CapitalG): robotics is a systems business, and the physical world has almost nothing you can buy off the shelf. Going full stack is complex, but it leaves you with a moat to compound on.

@_RobToews (Radical): some integration is required today because nothing is plug and play. But building your own hardware, pre-training your own models and running a service at once is like 12 companies in 1.

@ajay_bcv (BCV): it's hard to separate a robot's intelligence from its body. He suspects model companies are going to have their own embodiments, just like all the embodiment companies are trying to build their own model.

[Fellows Forum @Fellows_Forum]

.@wuseokjung_ on Corgi’s advantages over traditional underwriters:

"With the emergence of LLMs you can process large documents within seconds. Normally underwriters at companies like Chubb or Geico, it probably takes tens of thousands of people to go through these contracts.”

[Fellows Forum @Fellows_Forum]

.@pirroh (Replit) gave us some bold predictions during his fireside w/ @SriViswan at Fellows Forum.

Here's the full chat: https://t.co/9tUJJ8tdsd