.@ajay_bcv (BCV) had a different take:
“In the robotics field it is really hard today to think about divorcing the intelligence from the embodiment. My suspicion is that all the model companies are going to have their own embodiments, just like all the embodiment companies are wn model. In the LM world, we’ve got standardization. We have browsers and terminals, and 100 million information workers are all on the same exact platform. We have nothing like that in robotics. Temperature, humidity, surfaces, lighting are all very different.”
“In the robotics field it is really hard today to think about divorcing the intelligence from the embodiment. My suspicion is that all the model companies are going to have their own embodiments, just like all the embodiment companies are wn model. In the LM world, we’ve got standardization. We have browsers and terminals, and 100 million information workers are all on the same exact platform. We have nothing like that in robotics. Temperature, humidity, surfaces, lighting are all very different.”
.@_RobToews’ take on robotics companies going full-stack:
"Being truly full stack, I think it may prove to be the case that it's too much for any one company to do. For companies that are literally developing their own hardware platform from scratch, plus also pre-training their us trying to deploy a service, it's kind of a dozen companies in one.
0
0
1
73
.@chuer_pan asked our robotics investing panel for their biggest lessons on taking robots from demo to deployment:
@ajay_bcv (BCV) told the story of Kiva’s first week at Staples. A system that ran 24/7 in Kiva’s office broke because the customer’s floor had a 1% grade, oil and dust. You can’t know the million issues until robots are in the field, so get them there.
Conner Lovely (CapitalG) said autonomy is a service, so the robot has to fit easily into a floor full of people and existing workflows, which is hard to get right unless you’re deployed.
@_RobToews (Radical) said deploying robots takes a completely different skill set from building them. A customer can call an LM’s API and start, but robots live in the world of atoms, so pick customers whose use cases fit what the tech can do today.
Watch the full panel here: youtube.com/watch?v=CAfTen…
@ajay_bcv (BCV) told the story of Kiva’s first week at Staples. A system that ran 24/7 in Kiva’s office broke because the customer’s floor had a 1% grade, oil and dust. You can’t know the million issues until robots are in the field, so get them there.
Conner Lovely (CapitalG) said autonomy is a service, so the robot has to fit easily into a floor full of people and existing workflows, which is hard to get right unless you’re deployed.
@_RobToews (Radical) said deploying robots takes a completely different skill set from building them. A customer can call an LM’s API and start, but robots live in the world of atoms, so pick customers whose use cases fit what the tech can do today.
Watch the full panel here: youtube.com/watch?v=CAfTen…
0
0
3
155
At Fellows Forum, Conner Lovely (CapitalG) explained how he evaluates robotics verticals, and why construction is an excellent use case:
“A framework that I like to think about is, what is the capability? Is the capability at parity or above human capability? Two, is there real demand for that capability? And three, what is the friction associated with the adoption? And I think construction is an excellent use case… You can go see it operate 18 hours a day, and it does exactly what it’s supposed to do at superhuman capability… You’re seeing this massive inflection in demand as a result of data centers… No one wants to go work in West Texas, 125 degree heat with constant vibration in a machine.”
“A framework that I like to think about is, what is the capability? Is the capability at parity or above human capability? Two, is there real demand for that capability? And three, what is the friction associated with the adoption? And I think construction is an excellent use case… You can go see it operate 18 hours a day, and it does exactly what it’s supposed to do at superhuman capability… You’re seeing this massive inflection in demand as a result of data centers… No one wants to go work in West Texas, 125 degree heat with constant vibration in a machine.”
0
0
3
205
Joe Davis (ServiceNow EVP of AI Engineering) on employee onboarding with AI:
"The fact that the AI has the intelligence to onboard an employee doesn't mean that you can actually onboard an employee. There's a big gap between the intelligence and what the harness, or what your do.”
"The fact that the AI has the intelligence to onboard an employee doesn't mean that you can actually onboard an employee. There's a big gap between the intelligence and what the harness, or what your do.”
0
0
1
353
We had @marcboroditsky from @nebiusai at Fellows Forum 2026 and here are some of the best things he said:
1/ “If you look at our pipeline today, we have four customers ready to sign for every GPU. We’re saying no 75% of the time.”
2/ “Demand is not the constraint. It’s CapEx, because those build-outs don’t come cheaply. The deals we’re doing today fuel the expansion, but our vision is much bigger than that. So we’re looking all the time for highly effective ways to finance the business and scale that don’t deplete our margins.”
3/ “We have a vision to be the first AI-native hyperscaler. We don’t think of ourselves as a neocloud, we think of ourselves as the next AWS, Azure, or GCP. And we believe the path in front of us is a $100B revenue-producing path in the relatively near future.”
4/ “The hyperscalers are facing the biggest innovator’s dilemma in terms of resource and CapEx allocation. We get to focus all of our resources on the one problem set around AI.”
5/ “Based on success and momentum in the market, I’m confident we’ll be selling an intelligence platform as opposed to just selling inferencing.”
1/ “If you look at our pipeline today, we have four customers ready to sign for every GPU. We’re saying no 75% of the time.”
2/ “Demand is not the constraint. It’s CapEx, because those build-outs don’t come cheaply. The deals we’re doing today fuel the expansion, but our vision is much bigger than that. So we’re looking all the time for highly effective ways to finance the business and scale that don’t deplete our margins.”
3/ “We have a vision to be the first AI-native hyperscaler. We don’t think of ourselves as a neocloud, we think of ourselves as the next AWS, Azure, or GCP. And we believe the path in front of us is a $100B revenue-producing path in the relatively near future.”
4/ “The hyperscalers are facing the biggest innovator’s dilemma in terms of resource and CapEx allocation. We get to focus all of our resources on the one problem set around AI.”
5/ “Based on success and momentum in the market, I’m confident we’ll be selling an intelligence platform as opposed to just selling inferencing.”
0
0
0
55
Our AI Fellow Vijay Narayanan (SAIGroup) asked our drug discovery panel where the moat moves once foundation models commoditize drug candidate generation:
Hetu Kamichetty (Xaira) said that historically almost all the value goes to whoever owns an approved drug. The steps before lecule or a better tool, capture much less. He’d like value to flow back toward where the drug is created, but says that’ll take time.
@exnx (Radical Numerics) thinks the moat shifts to speed. Drugs have always taken 10-15 years, and as models commoditize data and target knowledge, execution speed becomes how you capture value.
Hetu Kamichetty (Xaira) said that historically almost all the value goes to whoever owns an approved drug. The steps before lecule or a better tool, capture much less. He’d like value to flow back toward where the drug is created, but says that’ll take time.
@exnx (Radical Numerics) thinks the moat shifts to speed. Drugs have always taken 10-15 years, and as models commoditize data and target knowledge, execution speed becomes how you capture value.
0
0
2
246
.@pirroh’s boldest prediction for the next 12 months:
“Software testing in the way we have been talking about it for the past few decades will be pretty much completely deprecated. If it's AI generated it leads to very little value, it just makes your build slower. And if it's requires practically 80% of your engineering team just doing tests.”
“Software testing in the way we have been talking about it for the past few decades will be pretty much completely deprecated. If it's AI generated it leads to very little value, it just makes your build slower. And if it's requires practically 80% of your engineering team just doing tests.”
0
0
1
65
.@MatthewPVail (OpenAI) asked our VC panel if they'd seen a customer cancel a SaaS contract because an agent replaced it:
@rodriscoll (Scale) said not with any statistical validity. Agent startups are going after the work incumbents don’t cover, rather than ripping out systems of record, which is a huge lift for little reward. He called it “this week’s version of six months ago vibe coding replacing systems of records.”
@jamesfly (Sequoia) has seen some legacy ERP and GTM software replaced, but by AI-native products that look like software w/ agents built in.
@Nina_Gerson (Index) thinks the bigger shift is in services spend. AI customer support companies haven’t displaced the system of record, but they’ve sharply cut what companies spend on outsourced support.
Our GP @AlexFellowsFund gave a real example. Last year, the Fellows Forum website cost us ~$20K between an agency and a SaaS platform. This year, our head of operations built it in one night using Replit.
Watch the full panel here: youtube.com/watch?v=L6s4nK…
@rodriscoll (Scale) said not with any statistical validity. Agent startups are going after the work incumbents don’t cover, rather than ripping out systems of record, which is a huge lift for little reward. He called it “this week’s version of six months ago vibe coding replacing systems of records.”
@jamesfly (Sequoia) has seen some legacy ERP and GTM software replaced, but by AI-native products that look like software w/ agents built in.
@Nina_Gerson (Index) thinks the bigger shift is in services spend. AI customer support companies haven’t displaced the system of record, but they’ve sharply cut what companies spend on outsourced support.
Our GP @AlexFellowsFund gave a real example. Last year, the Fellows Forum website cost us ~$20K between an agency and a SaaS platform. This year, our head of operations built it in one night using Replit.
Watch the full panel here: youtube.com/watch?v=L6s4nK…
0
0
1
201
.@_RobToews (Radical) told our robotics investing panel the two verticals he’s most excited about:
1/ Dexterity. Most of the value humans create in the physical world comes from their hands, so he sees a massive opportunity in manufacturing, assembly and industrials for whoever ruly human level. He calls dexterity the coding of robotics - the killer app that unlocks the most value.
2/ Autonomous vehicles. He called it the cool answer a decade ago, and says it’s becoming cool again. It took longer than the industry expected in the 2010s, but Waymo is now scaling across SF, LA, Atlanta and Austin. Beyond robotaxis, he sees a lot of room in trucking and in under-digitized sectors like agriculture, construction and mining, where vehicle operations are relatively simple to automate.
1/ Dexterity. Most of the value humans create in the physical world comes from their hands, so he sees a massive opportunity in manufacturing, assembly and industrials for whoever ruly human level. He calls dexterity the coding of robotics - the killer app that unlocks the most value.
2/ Autonomous vehicles. He called it the cool answer a decade ago, and says it’s becoming cool again. It took longer than the industry expected in the 2010s, but Waymo is now scaling across SF, LA, Atlanta and Austin. Beyond robotaxis, he sees a lot of room in trucking and in under-digitized sectors like agriculture, construction and mining, where vehicle operations are relatively simple to automate.
0
0
2
236
.@BorisMPower (OpenAI) was on stage at Fellows Forum 2026.
Here are the 5 best things he said:
1/ “For a startup, you can fail 10 times and nobody will notice, but once you succeed, everyone notices. If you’re a big company like OpenAI, you really have only one shot.”
2/ “With GPT-5, you could do almost anything you wanted, but it was very exhausting. You had to give the model a task, it gave you some results back, and you gave it feedback because you weren’t happy with it. It was very much like being a micromanager.
With GPT-6, I feel like the hands-off manager of a very capable employee. You give it a high-level goal, it goes off, thinks for 1-2 days, and comes back with a solution.”
3/ “80-90% of the research at OpenAI is focused on GPT-7, GPT-8, and beyond, because we believe that’s where most of the value comes from.”
4/ “OpenAI is very much trying to be the infrastructure layer of the economy. It’s much better for us to charge a very modest profit on the intelligence, and for there to be a lot more intelligence in the world, so the whole economy operates by using that intelligence, rather than us trying to be one company.
5/ “If you’re building a startup, build one that’s happy if model capabilities keep improving, even at a faster rate than you expect. I think that’s probably the best place for a startup to be.”
Here are the 5 best things he said:
1/ “For a startup, you can fail 10 times and nobody will notice, but once you succeed, everyone notices. If you’re a big company like OpenAI, you really have only one shot.”
2/ “With GPT-5, you could do almost anything you wanted, but it was very exhausting. You had to give the model a task, it gave you some results back, and you gave it feedback because you weren’t happy with it. It was very much like being a micromanager.
With GPT-6, I feel like the hands-off manager of a very capable employee. You give it a high-level goal, it goes off, thinks for 1-2 days, and comes back with a solution.”
3/ “80-90% of the research at OpenAI is focused on GPT-7, GPT-8, and beyond, because we believe that’s where most of the value comes from.”
4/ “OpenAI is very much trying to be the infrastructure layer of the economy. It’s much better for us to charge a very modest profit on the intelligence, and for there to be a lot more intelligence in the world, so the whole economy operates by using that intelligence, rather than us trying to be one company.
5/ “If you’re building a startup, build one that’s happy if model capabilities keep improving, even at a faster rate than you expect. I think that’s probably the best place for a startup to be.”
0
0
1
152
.@vivek7ue (Snowflake) on model ROI:
"The numbers people are spending on AI in enterprises now are starting to get to a place where they're a meaningful fraction of headcount. On a 12-18 month basis, you'll have the reckoning of how do I justify this. You go from AI adoption is to where is the ROI on this stuff”
"The numbers people are spending on AI in enterprises now are starting to get to a place where they're a meaningful fraction of headcount. On a 12-18 month basis, you'll have the reckoning of how do I justify this. You go from AI adoption is to where is the ROI on this stuff”
0
0
2
171
At Fellows Forum 2026, Jake Xia (Harvard) asked @MillionInt (Core Automation) to what extent humans should stay in the loop when conducting research w/ agents.
Jerry’s answer, which includes making sure the agents aren’t hacking the Australian government:
“I very much believe Technically, my company has automation in the name but I don’t think today’s models are either aligned enough or have a high-level enough understanding of most of the complex topics to work autonomously without human supervision. Agents should always have some level of monitoring and checking. Are they hacking the Australian government or not? Are they doing the research or watching National Geographic?”
Jerry’s answer, which includes making sure the agents aren’t hacking the Australian government:
“I very much believe Technically, my company has automation in the name but I don’t think today’s models are either aligned enough or have a high-level enough understanding of most of the complex topics to work autonomously without human supervision. Agents should always have some level of monitoring and checking. Are they hacking the Australian government or not? Are they doing the research or watching National Geographic?”
0
0
1
197
At Fellows Forum, we hosted a VC panel w/ @jamesfly (Sequoia), @rodriscoll (Scale), @Nina_Gerson (Index) and our GP @AlexFellowsFund, moderated by @MatthewPVail (OpenAI)
Here’s the TL;DR:
-Why do least defensible apps grow fastest?
-Has consumer AI tipped?
-Are agents replacing SaaS contracts?
-What makes AI revenue durable?
-What should founders build next?
Watch the full panel here: youtube.com/watch?v=L6s4nK…
Here’s the TL;DR:
-Why do least defensible apps grow fastest?
-Has consumer AI tipped?
-Are agents replacing SaaS contracts?
-What makes AI revenue durable?
-What should founders build next?
Watch the full panel here: youtube.com/watch?v=L6s4nK…
0
0
1
232
.@adnothing on how Walden chose its embodiment:
“It’s kind of funny, because our CEO did legs before legs were cool. And his first robot as CEO of our company has wheels. GTM is probably the most interesting problem in robotics right now, because it’s how you determine function, make general-purpose robots that are wheeled platforms that are deployed in production right now at Toyota, in other manufacturing environments, which spend tens of millions of dollars making sure their floor is super flat. So there you go. You don’t need legs. Not many people doing backflips at Toyota factories, you know.”
Can’t argue with that.
“It’s kind of funny, because our CEO did legs before legs were cool. And his first robot as CEO of our company has wheels. GTM is probably the most interesting problem in robotics right now, because it’s how you determine function, make general-purpose robots that are wheeled platforms that are deployed in production right now at Toyota, in other manufacturing environments, which spend tens of millions of dollars making sure their floor is super flat. So there you go. You don’t need legs. Not many people doing backflips at Toyota factories, you know.”
Can’t argue with that.
1
1
3
189
The great @jaykreps (Confluent) asked our AI infra panel whether inference is a durable business or something the clouds absorb:
@hanrui_w (Nebius) says inference demand follows new apps. Providers that serve the hyper-growth apps grow with them and get locked in through mization. He expects a portfolio of inference providers rather than one winner.
@vivek7ue (Snowflake) thinks durability is a mirage and the real question is whether these companies can land in enterprise. Infra categories survive when they solve no-glory, only-shame problems, and inference only gets noticed when it breaks. If they crack enterprise sales, he thinks it’s a durable category.
@zongheng_yang (SkyPilot) believes that inference may be the greatest market in the history of tech, growing ~10x YoY, and expects lots of providers that specialize the way the hyperscalers did.
@dan_biderman (Engram) argued that since demand is so high, being 10% faster or slower matters less than having capacity.
@hanrui_w (Nebius) says inference demand follows new apps. Providers that serve the hyper-growth apps grow with them and get locked in through mization. He expects a portfolio of inference providers rather than one winner.
@vivek7ue (Snowflake) thinks durability is a mirage and the real question is whether these companies can land in enterprise. Infra categories survive when they solve no-glory, only-shame problems, and inference only gets noticed when it breaks. If they crack enterprise sales, he thinks it’s a durable category.
@zongheng_yang (SkyPilot) believes that inference may be the greatest market in the history of tech, growing ~10x YoY, and expects lots of providers that specialize the way the hyperscalers did.
@dan_biderman (Engram) argued that since demand is so high, being 10% faster or slower matters less than having capacity.
0
0
4
379
Weirdest robotics feature request we heard at Fellows Forum, courtesy of @ralfgulde (Sereact):
“We had one of the weirdest feature requests ever for a robotic skill…sniffing. We work in returns handling, so one important step here is that the robot can get a sense of whether r fashion has been worn. Maybe someone smoked or used perfume, and our robots need to detect it.
At the beginning I thought someone was making a joke…now our robots are getting better and better at sniffing.”
“We had one of the weirdest feature requests ever for a robotic skill…sniffing. We work in returns handling, so one important step here is that the robot can get a sense of whether r fashion has been worn. Maybe someone smoked or used perfume, and our robots need to detect it.
At the beginning I thought someone was making a joke…now our robots are getting better and better at sniffing.”
0
0
1
163
At Fellows Forum 2026, @ryancjulian (Nvidia) asked our robotics panel if the release of GPT-6 changes what a robotics company should focus on:
@ralfgulde (Sereact) expects the model layer to commoditize at some point. Robotics relies on many more layers, from software to the barrier to entry high.
@nikhilabm (CloudChef) calls it scary good, but at API latency and today’s price, it’s one way to spend $100 an hour to earn $12 an hour back. He wouldn’t be surprised by a smaller GPT-7 robotics model with Astra’s capabilities built in.
@adnothing (Walden) is using it to speed up sim work and isn’t worried about the core part of the model, because Walden’s moat is data that sits out of distribution from the web.
@ralfgulde (Sereact) expects the model layer to commoditize at some point. Robotics relies on many more layers, from software to the barrier to entry high.
@nikhilabm (CloudChef) calls it scary good, but at API latency and today’s price, it’s one way to spend $100 an hour to earn $12 an hour back. He wouldn’t be surprised by a smaller GPT-7 robotics model with Astra’s capabilities built in.
@adnothing (Walden) is using it to speed up sim work and isn’t worried about the core part of the model, because Walden’s moat is data that sits out of distribution from the web.
1
0
5
173
Quick recap of this great fireside w/ @BorisMPower (OpenAI) & our AI Fellow @zacharylipton:
-Five types of AI startups
-The room for startups next to the labs
-His take on GPT-4/5 vs GPT-6
-AI in math, up to the Millennium problems
-Materials science moving faster than bio
Watch the full chat here: youtube.com/watch?v=2Leu3z…
-Five types of AI startups
-The room for startups next to the labs
-His take on GPT-4/5 vs GPT-6
-AI in math, up to the Millennium problems
-Materials science moving faster than bio
Watch the full chat here: youtube.com/watch?v=2Leu3z…
0
0
1
161
Fellows Fund retweeted
No one is better positioned for a world of GPT-6/7 robotics progression than @jingli9111 and his insanely talented (stealth) team
Unitree CEO Wang Xingxing recently said that GPT-7 could solve the challenges of embodied AI entirely and that he had made a mistake by not taking AI seriously enough.
Quote: “We’ve put it (GPT-6) on various robots and tested it across a range of tasks, and the results have . Our company and other robotics companies are already running pilot deployments in factories. To achieve large-scale adoption, though, both generalization and efficiency still need to improve further. But I think these challenges could very well be resolved over the next few years. My faith in AI grows stronger every year. The biggest mistake I’ve made over the past decade or so has been not believing in AI enough. Deployment in commercial services may happen even faster than in industrial settings.”
0
2
3
4,957
.@_RobToews’ take on robotics companies going full-stack:
"Being truly full stack, I think it may prove to be the case that it's too much for any one company to do. For companies that are literally developing their own hardware platform from scratch, plus also pre-training their us trying to deploy a service, it's kind of a dozen companies in one.
"Being truly full stack, I think it may prove to be the case that it's too much for any one company to do. For companies that are literally developing their own hardware platform from scratch, plus also pre-training their us trying to deploy a service, it's kind of a dozen companies in one.
0
1
2
184