Suhail Doshi

If your robot is not absolutely *beat up* from dogfooded usage, it's probably fake.

Best companies at model routing?

Dom is byfar the most innovative consumer thinker we have in AI. This is like some Twitch meets Vibe coded app/game thing?

.

Steadily knowledge workers are becoming Instruction Workers

🎯

Looking for 64 B200s with IB. DM.

The “AI cycle” roughly every 2-3 months I’ve observed for the past few years: 1. We want 1 frontier lab to prove they’re significantly better than everyone (AI progress is alive and well) 2. We want everyone else to nominally catch up (AI costs will not remain high, less centralization risk) [today] 3. We want OSS to exist to nominally catch up to #2 to provide an exit plan against the oligopoly / escape hatch (control performance/cost for your workload via finetuning) And as a friend cynically put it: “4. And then somewhere in there ______ drops the ball for a few quarters” (entertainment, chaos)

If anyone has gotten OPSD to work stably, could I ask you some questions?

That you can ask AI "why, why, why" as many times as you want like a 5-year old is a modern marvel of this century.

RT Paul Klein IV If I could only follow one person on X, it would be @levie, CEO of Box. We sat down to speak about why the REAL economy hasn't felt AI yet, the "model overhang" between capability and adoption, and what founders should actually build now that software is cheap. 00:00 - Cold open: agents use data like people do 01:53 - Is Redwood City the new AI epicenter? 05:18 - 90% of enterprise data is unstructured and was never automatable 07:46 - The model overhang: capability is outpacing adoption 10:35 - Why coding got the fastest AI takeoff 16:24 - Inside Box's agent-first rebuild 19:15 - Headless SaaS: agents will outnumber humans 100:1 25:41 - Model routing + why the applied layer wins, even bitter-lesson-pilled 27:45 - GTM is the new moat Checkout Navigators, our new pod on Youtube, Spotify, and Apple Podcasts!

RT Suhail Looking for 64 B200s. DM if you know a vendor.

RT Keith Rabois As predicted….

RT Ti Morse My third interview with @Isaiah_p_Taylor, Founder of @ValarAtomics. Valar just became the first startup in history to power an NVIDIA Blackwell with a nuclear reactor. 0:32 Manufacturing nuclear reactors is a SpaceX style problem 2:58 Scaling from 1 reactor to 10 4:31 Why it’s so tempting to iterate on paper instead of actually building reactors 7:37 Having contact with reality - modular shielding 10:09 Steel is cheaper than software engineers 11:09 Why the idiot index is so high in nuclear 16:02 Becoming a nuclear company 18:57 It's important not to lie to yourself 21:55 Obsessing over critical path 24:54 Nuclear reactors are relatively simple 30:35 Safety is a function of iteration speed 30:55 Why Valar had to vertically integrate 35:01 Designing a Toyota Camry vs a Ferrari 38:07 Running through 1-way doors 41:13 Pulling rabbits out of a hat 46:13 Wartime mode 48:49 Being willing to do unreasonable things 51:20 How Elon injects urgency into his companies 56:56 Relentlessly chasing the goal of making energy 10x cheaper 58:32 What’s changed over the last 3yrs 1:00:51 Designing safe nuclear reactors 1:13:05 Learning from SpaceX - optimizing for scale 1:18:07 Having a high tolerance for looking dumb 1:19:50 How Palmer Luckey handles hit pieces 1:20:55 Inevitable time 1:25:26 Predicting unknown bottlenecks 1:28:08 Scaling 1:29:45 July 4th will be a rebirth for nuclear in the United States

Cost per task is going to be a meaningful metric these next 12 months.

Re Some very cool people in the DMs. Thank you - will get to you soon. Caught a cold.

The raw reasoning traces of these models is pretty funny. Most people don’t see them since they’re hidden on the closed models. They constantly overthink, kind of like us in our minds.

Excellent.

The mass construction of new datacenters and power across the US reminds me of the mass buildout of the Internet. Whenever something like this happens, something new becomes true that previously everyone concluded wasn’t possible and have written off as a good idea.

Beware: "The most durable way to do that is not to out-compete the open frontier but to regulate it"

This is a very good entry into post training LLMs with RL. The whole recipe and data is open. Highly recommend!

/goal for AI model training runs is *so* good - it really feels like the future. Very little babysitting now. Mine: Launch a full training run on 4 nodes. Continuously record things in an experiment document if it exists. Log hyper params, configs, periodic evals, performance insights, analyze training stability, and important changes for future analysis and reproducibility. Fix any major bugs you encounter while you monitor training but do not change the fundamental nature of the experiment without asking. If it crashes, resume again and keep training. Resume from latest reliable checkpoint you have. Reach steps

I've started doing a lot of async learning where I'll take every interesting blog post, tweet, arxiv paper and ask AI to teach me it with a specific prompt. I never read it in the moment, I just queue it up for later. Then when I have a block of time (in an uber, before bed, etc), I'll go and read it vs doomscrolling X. This is somehow become better than saving links in Notes or something because when I finally have time to read it, there's this highly approachable artifact for me to dive into, read, and ask follow up questions. I feel like I am learning at a much faster rate than I ever have before. Even if the content was poorly written, AI can fix it to your liking and adapt to your experience level.

PSA in SF: it’s July 4 soon so random illegal firework testing is happening if you’re wondering if a bomb went off.

Well, yeah.

It's that time of the night boys.

Every time I acquire more compute, I discover that I could use even more compute to do other new things.

Very excited about this!

Which means, somebody will make a lot of money commoditizing it vs charging a subscription for it.

Nowadays it seems like you’d have to be actively working on the frontier to be prepared for the moment of something in your lane. It’s less clear to me that you can calmly observe, assume you have 6 mo to build, and then launch. You’d have to be building something adjacent now.

RT Isaiah Taylor - making nuclear reactors Good morning. Last night at 8:51pm Mountain Time, the Valar Atomics Nuclear Operations team brought Ward 250 to 10kW thermal output. We held this power level throughout the night. This is another historic milestone from the Valar team, marking first nuclear power by a startup.

Happy Father’s Day! My kid ran into our bedroom this morning swinging a green lightsaber yelling “You are Darth Vader!” at 6am. He also brought a card ❤️.

The next 10 days of my life do not have a single meeting or external obligation. The glorious freedom to focus.

🎯

Everyone is cycling through all the labs. Excited for trade secrets to get recycled.

RT Isaiah Taylor - making nuclear reactors Moments ago, Valar Atomics took Ward 250 critical for the first time. This fulfills President Trump’s EO 14301, which called for 3 advanced reactors to go critical by July 4th. This is our second criticality as a company, and an important step toward our goal of power by July 4.

I’ve heard about rumors of this for over 12 months. @DavidSHolz has a level of ingenuity and original thinking rarely seen.

RT Oliver Cameron When @jeffrey_hawke and I started @odysseyml in 2023, we believed general world models would become a new class of foundation model. Three years later, it now feels true. And today, we're announcing a $310M Series B. AI can now understand and simulate the world!

This week will quickly demonstrate how much anyone will trust Anthropic around these doomsday / natsec claims. You can only cry wolf so many times.

RT David Sacks I’ve had a number of conversations with folks inside and outside government about the current situation with Anthropic, and here is what I believe to be true: — As we know, Anthropic publicly released its Mythos class models earlier this week under the commercial name Fable. — Fable is Mythos with guardrails. But if those guardrails fail, then you’ve exposed Mythos and its advanced cyber capabilities to people who shouldn’t have them. (Keep in mind that Anthropic itself widely promoted the idea that Mythos was a cyberweapon and needed to be regulated as such. They asked for government regulation of Mythos and championed the guardrails on Fable. If there is a vulnerability — big or small — it is Anthropic’s responsibility to patch.) — A highly credible trusted partner of both Anthropic and the USG who was testing Fable came forward with a jailbreak of those guardrails. The Admin asked Dario to fix the jailbreak or de-deploy the model. Dario refused. — In their blog post, Anthropic defended its decision by saying the jailbreak isn’t serious. That is not what the trusted partner and the USG believe; nor is that kind of minimizing language consistent with Anthropic’s brand as the AI safety company. It’s difficult to fathom how they could claim a jailbreak allowing operability of a cyber weapon could be defined as not “serious.” — In the past, Anthropic has always said that safety must be top priority and taken super seriously. In this case, Anthropic prioritized the continued offering of the consumer model over safety. — In reaction, the Admin issued the export control. The Admin did this reluctantly. It’s been very surprised that Anthropic hasn’t wanted to cooperate with a reasonable safety request (ie fixing the jailbreak issue). Anthropic’s reaction is very much at odds with their branding and ethos as a safe AI research community. — The Admin’s hope now is that Anthropic remediates the safety issue, the export control is lifted, and Fable ...

It’s naive to think anything is backfiring here. This is part of the expected plan for reg capture. Of course the govt steps in such that you design and define the rules you already follow but are now required by your rivals. For their AI could be too dangerous.

The end-game for Anthropic is becoming government controlled by a single nation. As Thiel once said: Going IPO is like a government takeover (quasi empowering CFO, lawyers, etc as govt actors). Regulatory capture for capability oversight is a step towards monopoly.

My favorite @elonmusk quote that I often send friends: Do not fear losing. “You will lose,” Musk says. “It will hurt the first fifty times. When you get used to losing, you will play each game with less emotion.” You will be more fearless, take more risks.

SpaceX feels like a bet on rooting for humanity's grand ability to invent, thrive, and far surpass our wildest imaginations. Bravo to you all that made this happen.

Priorities for high agency people are almost always communicated by the latency of their response.

Re Lots of great content out there but sometimes the writing sucks or it’s written for a different audience that has more knowledge than you. AI is a great at re-interpretation.

Met both my competitors: one where we beat them and another where they beat us in a segment. I think the most helpful thing to help you grow is to ask your competitor what they think you failed at when the time is right. Generally they have great insights. Cost is your ego.

> “AI is progressing extremely fast—much faster than the policy process was built to handle” It’s not evident to me that any policy process is needed at all.

RT martin_casado Imagine building a computer and not allowing its use in CS research. Thats some dystopian shit.

Yep 🎯

Users really dislike censorship but they really really dislike invisible censorship. Brand damage is going to be unrelenting here. Lack of sufficient competition makes companies do surprising things until they no longer can afford to.

RT Sriram Krishnan just to state the obvious: think there's a collison course between those who believe research and science should be open and those who believe we are in an accelerating singularity curve. I have many smart friends who have believed both for a while but seeing more and more their realization that these beliefs will be in conflict. I for one believe that America and the west needs open and distributed access to research and computation and sharing of ideas at all times.

Very true. Often w startups there's a near infinite surface to critique. However, it takes very little to keep the flame flickering just a little bit longer for relatively formidable people to breakthrough to other side.

Newly concerning for the AGI optimists...

Having spent a considerable amount of energy studying the robotics world, this is 🎯

Beware of the bio weapon that could result from this 👇

> What does the D stand for in DNA? Refusal!

I would like to +1 that this is a very bad policy. Respond with a refusal and deal with the fall out but invisible NERFing is super uncool.

RT Saurav Shroff if you are on Transporter 18, I want to fly a coupon of a next-gen product on your vehicle I'll pay a minimum of 100k, maximum of 1M, and offer a minimum of 100k, maximum of 1M, in free credits for future products from @StarlightSolar_

It is incredible to watch prices go up for American AI models. Like the frog being boiled with each update.

I had not fully considered this possibility before. Interesting.

Ah the peace of being back home to work and build things in a quiet space 👌🏾

Blue pill or red pill?

Step 1: Do a shit ton of inference Step 2: ? Step 3: Profit

All my bad VC stories mostly just make me sound like a wuss so I'll just share a good one: One time I crashed an Allen&Co event since it meant I could pitch 4 investors in one day in the same location. I didn't want a week gap in my fundraise so Max Levchin encouraged me to crash the event in Scottsdale after I joked about doing so. I flew to AZ and drove w my dad in the car to the Ritz. Anyway, everyone I pitched was very lukewarm until I get to Ben and Marc at a16z. I pitch them both at a coffee table. Neither seemed all that interested in my deck so I presumed they're also checked out on my company. Ready to close the laptop and return home, Marc stands up and says: "If anyone gives you an exploding termsheet, tell them to go fuck themselves." At this point, I hadn't even heard of what that was so I had no idea if this was a good or bad reaction. It was Friday. I went to the partner meeting on Monday. Termsheet that week. We had no other termsheet options. The rest is history. I really appreciate the conviction they had on two young nobodies.

man the early days of ai were so special :)

A reminder:

I have been building an RSI harness since about late November. It works fantastically well at a number of problems: top 18% in most Kaggle competitions, new techniques for ARC-AGI 2, discovers model optimizations. It mainly lacks creativity. Once we resolve that, RSI is near.

1/ it all started w 2 8xB200s excited to be back in the game again

A tremendous amount of startups is the acceptance of the risks you can't know anything about and the conviction of the future ahead. This is why you'll often meet people very skeptical and equally enthusiastic along the way.

It is astounding how much and how fast you can learn anything with LLMs. On one hand, you could devalue intelligence / sulk or you can just be some guy in a small room learning the absolute frontier of your field at any given hour. Self-teaching has never been so diffuse.

There are 15x engineers, 4x engineers, and 1x engineers. If you don’t use AI at all, you may be sub 1x. P99 at lines of code automatically places you in sub-1x.

11 is an even row window according to GPT 5.5 thinking.

A new era of Suhail. 25.0528, 121.5990

New roadmap for AGI. Maybe we should call it The Pope Test.

If anyone has a clue about how to handle skill obsolescence, it's Lee Sedol. He basically went meta on it.

The answer is to work on harder, more complex things. We must sharpen our skills for the next frontier.

RT Paul Buchheit What it takes to reach $100B Fewer than 0.1% of startups will ever be worth more than $100B, but those that do will have an outsized impact, so it’s worth understanding which companies have the potential and what it takes to get there. Examining the history of these massively successful companies, it becomes clear that there are two ingredients necessary to reach $100B. First, they must be building in a rapidly growing market of unlimited size. For example, Microsoft, Apple, Intel and AMD all emerged as part of the exponentially growing microcomputer market. These companies started when microcomputers were still relatively new and obscure. Micro-soft’s first product, Altair BASIC, was incredibly niche — MITS only ever sold about 25,000 Altairs, but that was the start of what is now a $3T company. Likewise, Amazon, Google, and Facebook all became $T companies by growing with the Internet. Stripe ($160B) makes this dynamic explicit in its mission statement: “Our mission is to increase the GDP of the internet”. Why now? $100B opportunities only exist for a limited time. If a company could have been started 20 years earlier, then it’s unlikely to have $100B potential. Important new technologies create massive new opportunities, but those windows of opportunity don’t last forever. For example, it was not possible to start Uber or DoorDash five years earlier because mobile platforms such as the iPhone did not yet exist, and it wasn’t possible to create them five years later because the opportunity had already been captured. Large but slow growing markets rarely produce $100B companies. For example, startups selling to dentists or auto mechanics are not good candidates to reach $100B. A simple test is to ask if demand will increase 10x or 100x in the next ten years. Startups thrive when capturing a slice of a rapidly growing pie, not fighting zero-sum games against incumbents. The second ingredient is defensibility, a durable control point in the ma...

It feels like replies are going to have to move to a trust based system due to AI. If anyone can generate a plausible one, it doesn't seem to be valuable for the algo. Thus, I foresee only my follows and my follow's follows as the new default.

Hell of a run @drewhouston - congrats on *19 years*.

Possibly the thing we will most realize looking back: intelligence was so big that lots of companies were going to succeed. It's not so simply bucketed into chatgpt and claude code.

Nothing seems to replace understanding the thing you're working on if you intend to improve it. AI generating 2K lines of slop may initially work but once you want to tweak it or understand the nuance of what's happening under the hood, you end up wishing you had built it up.

🎯

😫

I think people classify things as a commodity very casually these days. From hardware to router/switches to cutting metals to AI training clusters. It's pretty lazy because if you spend a good amount of time, you'll often see how much skill and understanding is involved. If someone seems lazy in saying things are commodities, I'd be default skeptical. And if enough people say it, you should find out if they are right because it may be a big opportunity in disguise.

RT Garry Tan The companies I love working with in office hours are the ones where the founder has a specific, weird, earned insight that nobody else has. Not "AI for X." A genuine edge that came from living inside a problem. The ones that are dying almost always have the same pattern: technically competent founders building something nobody asked for, moving metrics that don't matter, avoiding the conversation with the one user who'd tell them the truth. The lucky thing is that 2nd type of founder can become the 1st kind if they don't stand still, they are willing to talk to people, try things, and always seek high rate of learning.

One real issue with Chinese independence of flop-based compute is that their open source contributions will shift to a stack we can’t / won’t use in the US and that will hurt because our current AI research / infra is currently very closed.

Future of how we will design.

Paraphrasing Jensen here: “I am glad they’re experimenting with other chips. How else would they know how good ours is?”

RT Aditya Agarwal 4 thoughts on early-stage hiring: 1/ If an engineer is trying to pick between a pre Series-B company and a BigCo/BigLab --> stop talking to them immediately. They are clearly not ready for a startup. 2/ If someone isn't willing to take a 70% cash paycut (relative to BigCo/BigLab) --> stop talking to them immediately. They will be unhappy/stressed. 3/ You learn a lot about a candidate during the negotiation/closing process. Do not be afraid to walk away if you get new information. 4/ Startups have zero work-life balance. If you are not willing to put in the hours, you are not in the right headspace to grind.

I had no idea Aliens would need B2B Enterprise SaaS too. Wild.

Just invest based on the pictures 🫡

RT Y Combinator Pops (@pops_fyi) lets anyone create and play AI games. Describe your idea, and Pops turns it into a fun mini game to remix, share, and play with friends. No code required. All you need is an idea. Short-form software. Congrats on the launch, @alonzuman! https://www.ycombinator.com/launches/QPl-pops-create-and-play-ai-games-with-friends

I remember sitting in coffee shops w Immad while he waited for approval to launch this thing. *4 years of profitability* Unreal.

An advantage often underestimated time and time again. 🎯

Holy cow. We have come a long way. Breakthroughs every single week, still!

I finally got around to the Jensen/Dwarkesh video. So good. But my favorite line was: "We do as much as we need to do as little as possible" in connection why they don't go into the datacenter biz which ofc extends to everything else. This is the opposite of how the labs think.

There’s this feeling you get like when you learn a new word and suddenly you hear it everywhere that happens with something you can become interested in. Suddenly, every link/story/tweet/person interaction ends up being around that thing so effortlessly. It’s everywhere you are.

Can’t even read the article and I know it’s about @isaiah_p_taylor

Ah, remember when knowing JavaScript was a skill. Those were the days.