Rory O'Driscoll

Wrote a blog post to expand on this post from a few days ago.

nextview.vc/blog/the-three…

Feels like every week in robotics there’s a new ‘this is the GPT-3 moment for robotics 🤖’ announcement. 

We brought on @GeneralistAI CEO & Co-Founder @peteflorence on @GreylockVC Change Agents to dig into what’s going on at the frontier of robotics models.

We covered the company’s latest Gen 1.5 model, few-shot learning, training robots on different embodiments, and the milestones towards a more generalized physical model. 

Timestamps:
00:51 The inception of Generalist
04:13 Long-term goal of the company
05:47 Parallels and differences between robotics and language models
12:05 Key differentiation in 1.5 Gen model
12:48 Robot vs banana
14:49 Emergent capabilities not explicitly trained for
17:13 Cross-embodiment and the importance of hands
22:26 Research vs working with customers
24:58 Beyond VLA vs world model
28:50 Future-looking milestones

Some of the top takeaways:

- One-shot and few-shot learning emerged without being trained for. Gen 1.5 can learn a new task from a single demonstration, and Pete compares it to the GPT-3 moment in language. In one example, a robot taught to sweep a cube into a bowl with a brush used a banana instead. Given a dustpan, it held the pan with one hand, swept with the other, then tipped the cube into the bowl. Neither behavior was explicitly trained, and Pete sees this as a signal of where the model's generalization is headed.

-Generalizing to new hands remains a challenging problem for cross-embodiment. Physical hardware doesn’t stay static, and so cross-embodiment - the ability of a physical AI model to adapt to different hardware systems - is vital for success. 

-Customer deployments are a valuable source of research inputs. Generalist actively partners with their customers for feedback, which they use to inform their research and make real-world evaluations.

-Generalist doesn't think in terms of "VLA vs. world model." Pete helped create early VLAs and has worked on world models, but he argues the goals matter more than the label, and the team is trained to think in a first-principled way when considering new research directions.

Watch the full episode at the link in the comments. Thank you to @peteflorence for joining us!

We’re excited to lead @namespacelabs's $42M Series B as they build the development platform for the agentic software engineering era.

Developers have always worked from a laptop holding their carefully accumulated environment, with CI turning the code into working software. Coding agents break that model, both by increasing the volume and concurrency a platform has to support and because an agent often needs a build or test result before it can decide what to do next. Every slow run delays the next iteration.

Namespace is a development platform designed to serve engineering teams and fleets of coding agents alike. Its devboxes give every agent an isolated workspace that starts fast and keeps useful state from one iteration to the next. Namespace also operates its own infrastructure, which lets it schedule work close to the dependencies and build artifacts it can reuse.

We’re especially proud to support CEO and Founder @20thr, who built a development platform from the ground up as a principal engineer at Google. Frontier AI teams like SpaceXAI and Ramp are already building on Namespace.

Read Partner Javier Redondo’s post on how Namespace is unlocking the new software development lifecycle: scalevp.com/blog/announcin…

Yesterday we previewed Griffin, a Human Interaction Model capable of seeing, hearing, sounding, and looking like a human does. There have been a lot of questions, so I wanted to take a moment to share our thoughts.

By way of introduction: Tavus is a research lab focused on enabling machines to meet us where we are, and to understand the nuances of how we communicate beyond words.

Griffin, our latest model, isn’t publicly available yet. Yesterday’s announcement was a limited research preview to demonstrate what the model is capable of. With AI progressing so quickly, we prefer to share breakthroughs openly and in real-time as we work on a safe public release.

Face-to-face is how we evolutionarily communicate, it carries the most meaning and intent, and we want computers to be able to help with work that benefits from that emotional understanding, expression, and immersion. Some examples of the kinds of use cases we care deeply about:

- A tutor that can build understanding of how a student learns, see exactly when there is confusion or disengagement, and adapt the lesson to fit them.

- A health expert that can answer any questions about your upcoming appointment or prescription, at the pace you want, at any time you need, even on a weekend.

- A language coach that you can practice speaking with, that can correct your movement and pronunciation, and help build confidence to have real conversations

- Or the perfect assistant for everyone, that understands intent, knows how you work and remembers what matters.

We’re working with partners on safeguards and systems for disclosure, as well as inviting discussions with officials around wider regulation and safe use. People will always know they’re interacting with AI, while providing an interface that removes the need to ‘speak computer’.

We believe in a future where computers understand us well enough to make technology more accessible, more useful, and make us more capable as humans. That is the world we want to build, and we understand the responsibility to do so safely.

This is the culmination of years of work from the @tavus team. The quality is simply amazing. This will enable a way more interactive, human-centered computing experience. Congrats @hassaanraza @quinnfavret and the team.

For some reason, every tech bull market seems to come with an apocalypse story. In 1999 it was Y2K. I remember New Year’s Eve in San Francisco, half the people were drunk because they were rich and the other half were up all night making sure the computers didn’t stop working. Today the money is bigger and the apocalypse is bigger. Now we’re not just going to lose the computers, we’re all going to die. PSA: January 1st, 2000, the computers worked just fine, and the money vanished three months later.

Exactly. At some point, a Congressional Committee is going to subpoena these CEOs and say either a) you think there is a 10% chance you are gonna blow up the world, in which case shut it down, or we will shut it down for you, or b) you think your own safety people are talking overwrought doomer BS, in which case shut down the rhetoric, you are scaring my constituents. It cannot continue like this.

There are four things that can go wrong for Nvidia, but only end-user demand really matters.

1. Direct customers stop buying compute. Not going to happen, the hyperscalers are exploding. 

2. Worries around roundtripping cause problems. Also not happening, Nvidia is kicking off so much cash that they can’t spend it fast enough.  

3. Competition. In theory, one or five other startups could take significant market share. It’s probably already happening, but when you start at 90%+, it takes a long time to show up in the numbers.

4. End-user demand. This is all that matters. Everybody gets to sell chips to hyperscalers, provided hyperscalers can sell compute to OpenAI and Anthropic, provided they can sell intelligence to end customers. Right now that’s happening, so everything down the line is plus or minus fine. If it ever unravels, it won’t be because of roundtripping. It will be because the frontier models forecasted 5x growth in end-user demand and only got 3x.

Lucky enough to have @DroneDeploy and @fin_ai close this week. Happy for both teams but genuinely sad to no longer work with @mikewinn and @eoghan. Two examples of savvy big company SaaS acquirers (@procoretech and @salesforce) leaning into AI with acquisitions that can meaningfully move the needle. Congrats to all.

Big acquisitions are driven by paranoia, but “just because you are paranoid doesn’t mean they’re not out to get you.” When the frontier models are claiming TAMs the size of US GDP, every other tech company has no choice but to respond.

The recent spate of big acquisitions (Hugging Face, Cursor, and OpenRouter) is about extracting value away from the frontier models, either by enabling the open weight ecosystem (Hugging Face and OpenRouter) or by taking the largest frontier model market: coding (Cursor) and building a viable competitor to Claude. They won’t be the last.

When the AI IPOs come, the stock based comp numbers are going to be nothing like we’ve ever seen. None of that shit will matter. The only thing that matters is the growth rate and the 27 and 28 projected revenue.

The reason you normally worry about SBC is because in a steady state business like Workday it really is a cash number. If you’re giving someone 500 grand every year to show up and be a middle manager, they are mentally putting those RSUs into their comp. If you stop giving them RSUs, they’re going to want cash. In a mature business, it’s totally correct to worry about SBC.

A hypergrowth AI company is not that. Someone hired with a million dollar package in 2023 ended up making 51M four years later. That doesn’t mean you have to pay the next guy 51M. He would have signed up for the million he was offered. That’s the real economic stock based comp. The other 50M is just dumb luck. So it’s okay in a hypergrowth company to look past a good slug of the SBC and normalize it out. And conversely it’s not okay in a mature company because that’s real money that people are spending.

It might be a little unfair that the hypergrowth company gets a free pass and the mature one doesn’t, but they get a free pass anyway, provided that revenue goes up. Once revenue stops going up, all bets are off.

The most important number in the market today is the OpenAI Q3 ARR number. The trailing numbers had GAAP revenue of $6.7B in Q2, growing 18% q/q. This would point to sub 100% Y/Y growth, rapid deceleration and GAAP revenue under $30B for 2026.

Amazing numbers in any normal world but a company with those numbers will not buy $750B of compute by 2030 as OpenAI has promised to do. The result would be “bad” for the AI trade. That is why the company is (wisely) leaking strong July and Aug reacceleration numbers and analysts are crawling through CFO comments like Kremlinologists of old trying to figure out what “35% quarter to date growth” means. Are we back to 3x plus Y/Y growth?

Companies are starting to think about next year's AI budgets, and the big question is what happens to overall spending. There are two forces at play.

1. There are maybe 5% of companies who tokenmaxxed earlier this year and are now “tokenminimizing” like crazy. Coinbase is the poster child here and is pointing to a 50% reduction in spend from the peak.

2. And then there are the toedippers, the 95% of companies who have just started on AI. If even 25% of those companies start to expand, their growth will swamp the reduction from the tokenmaxxers.

Tokenminimizers vs toedippers will be what decides FY27 budgets.

A platform shift is an ever expanding crevasse between the before and after. If you’re going to make that jump, you have to make it fast, because there comes a point where the gap gets too big to cross. Palantir saw it with AI in 2023 and went from 18% growth to 98%, meanwhile, many SaaS companies are only now nerving up to make the jump. There’s going to be a lot of companies paying the bill in 2026 and 2027 for the hesitancy they showed in 2023 and 2024.

Whatnot raising at a 20 billion valuation doing live shopping shouldn’t surprise anyone. Shopping + Fun equals big bucks. QVC and the home shopping network ran the same exact model on TV for decades, and people loved it. eBay auctions were the same thing online in the 90s and it’s still worth 45 billion today. There is more to life and venture than LLMs.

I don’t completely agree with Harry’s take around A-tier talent because it's all about what kind of A-tier talent you want. Nobody building application software companies in the PC era had A-tier chip talent, because they weren’t building chips. They were taking advantage of all those smart Intel engineers and Microsoft operating system engineers and building on top.

The same logic applies today. If you’re actually building a frontier model and you don’t have A-tier model talent, then yes, you’re toast. But if you’re not, you don’t need A-tier model building talent. You need A-tier talent in USING frontier models to build great apps.

Well done. I wish I had tweeted that @joshk .

Agree I should have been more precise in my terminology. The wider point is the old napoleon line about indispensable men.

You are welcome and congrats on the role . My best to scale from scale

I was totally wrong on Scale AI. I had originally thought that the acquisition and the exodus would leave them a husk.

It proves that when you're in a great market with a product that meets real demand, even losing your top people won't kill you. Congrats on the great CEO hire.