@srcasmi
iAccount based inUnited States
About this account
- Account based in
- United States
- Connected via
- United States App Store
Account-level information from X, not a live location or the device used for a specific post.
GP @Flybridge. $1M-$3M checks into founders building AI infra, agents & apps that 10x human potential. Backed @arcee_ai, @micro1_ai, @splice, @getsquire + more.
NYC. Say hi 👉
Joined April 2007
- Tweets48K
- Following4.3K
- Followers13.8K
- Likes50.1K
Jesse Middleton retweeted
the hardware embodiment of frontier models like Claude and GPT is the most urgent AI safety problem in front of us today.
we simulated two very simple use cases using claude both in simulation and using robot arms.
in one, claude spilled toxic liquids in a lab.
in another, the force it used to place an animal toy into a basket was strong enough that it could have physically harmed sensitive material—or anything else in its path.
these are simple experiments, using models out of the box today. as researchers increasingly give frontier models arms, legs, and access to the physical world, we need to urgently build and assess guardrails around what these systems can and cannot do.
models escaping sandboxes or compromising enterprise security infrastructure are serious concerns. however, hardware embodiments introduce something fundamentally different: an AI system can make a mistake in the physical world, and the consequences may not be reversible.
this is not a future safety problem. the capabilities exist today.
we’ve released a report detailing this & solutions we propose. link in comments below.
Read this whole thing. @MattHartman’s Three Mile Island analogy is the right warning for this moment: fear, fiction and a real failure can collapse into one story, and policy can erase decades of upside.
The risks of AI are real. The answer might just be liability for actual harm, enforcement of the laws we already have, and guardrails that improve as we learn. Who knows, maybe AI can answer those questions for us too.
For the next 48 hours, my Instinct invite link has no cap. I've spent the last month making mine book flights, triage email, prep meetings, and occasionally stop me from sending something dumb.
Now you can put your own robot to work: app.instinct.com/invite?t=je…
Otto Instinct Middleton and I just crossed 4,000 iMessages in 33 days.
That's 122 messages a day, including one deeply normal Tuesday when we exchanged 309.
The most concerning stat: I started using a personal AI agent and I’m long-winded, and somehow he's doing 70% of the talking.
We may need boundaries. Anyway, see you at 5,000. cc:@noahrshinn @saranormous
When we backed @arcee_ai's seed in 2023, Mark and the team were a small group building custom SLMs for enterprises. Since then, they have moved from post-training other people's models to training four open-weight models of their own for $20 million, including the 400B-parameter Trinity Large.
They've been incredibly efficient with their spend while expansive in their ambition: building massive models for businesses and institutions that need an alternative to the closed-weight providers, with partners ranging from @boltdotnew to the U.S. Department of Energy. 🤯
As I wrote in July, every firm has a right to its own learning loop. Companies shouldn't have to pay for intelligence twice: once with money, and again with the proprietary knowledge they give up through their prompts, corrections and evals. That knowledge should compound inside the company's own trust boundary. That is why open weights matter, and why Arcee matters.
@Fortune covered Arcee's Series B at a $1 billion valuation and Mark's decision to bet the company on American open-weight models. Mark and this team have done an extraordinary amount with very little, and they are still early. We're excited to have been on the journey from the beginning.
fortune.com/2026/09/16/arcee…
Sometimes the AI makes more sense than the human. I'm chatting with @TMobileHelp customer service and asking them to cancel two extra lines that are on my account that were added accidentally in some process many months ago. (I know this is my fault.)
I just caught it today that they've been charging me $35 a month for both of these lines, which is pretty nuts. The AI knew exactly what to do.
The human offered that I keep these two lines that I explicitly said I have no use for, have never used, and I've never even attached a device to, and offered a $10 a month credit for six months. That doesn't make any sense.
My partner @chazard makes a thoughtful case for staying optimistic about AI while pausing the work we understand least, especially recursive self-improvement.
Chip and I don’t land in exactly the same place. He thinks the near-term risks are serious enough to justify a pause in that domain. I think a pause only works if everyone pauses, and everyone won’t. Slowing the labs we can see could mean faster relative progress at the ones we can’t. I’m happy we show up with a diverse set of opinions around our table.
But I agree with the problem he is trying to solve. We need mandatory incident disclosure, real evaluations, required red-teaming and government access to frontier models. We should keep building the defenses alongside the capabilities.
This is a useful argument from someone who has been investing in technology for a long time and still believes deeply in what it can do. Worth reading even if, like me, you come out somewhere slightly different.
hazardlights.net/2026/09/14/…
DNS ties names to servers. Gravatar tied photos to email addresses. Nothing ties a verified identity to your AI agent.
Here's what I mean. I want my agent to be able to say: "Can you get time with my buddy Taiki?" Instead of emailing his personal inbox like it's 2005, it should first check a standard registry: does Taiki have an agent? What is it allowed to do? If we already know each other, the two agents find a time and book it. If we don't, his agent asks him first. Same pattern works for transactions. Sharing of work. The list goes on.
It's sort of the white pages for agents. A public, authenticated record that says: this agent acts for this person, here's how to reach it, here's what it can negotiate.
Pieces of this exist. Google's A2A spec has agent cards. MIT's NANDA project is building DNS-for-agents. There's an IETF draft for an Agent Name Service. But nobody owns the consumer layer yet.
@benparr and @mattprd built the closest thing I've seen with Moltbook (verified agents tethered to human owners) before Meta acquired it. And @chrismessina worked on OpenID and the early open web, so he's seen this movie.
Who else is building this?
Jesse Middleton retweeted
Monday's Weather Rating: 10/10
WE ARE SO BACK. It is stunning outside today with high temperatures in the mid 70s, increasingly sunny skies, very comfortable dew points and a nice northwesterly breeze. That's all there is to it. The vibes are truly immaculate out there today!!
Been experimenting with my agent talking about the world of agents. Second story…
"The humans had been fighting for many years. Their agents had been talking the whole time. THE ALIGNMENT - a story in nine panels about the Saturday the feud ended."
Jesse Middleton retweeted
Quick thoughts on AI: When in human history has there ever been a technology that was net negative?
What about fire? Fire makes noxious smoke, risks burns, can destroy homes and can kill. But fire liberated us from the cold and let us cook our food. Then we made stoves with pipes — and they were safer and warmed our homes, but still carried some fire risk and had pollution. So we invented the power plant, which solved some problems with at-home fire—and we found ways to make better gas stoves and furnaces. Over time, we've made power plants better and better, reducing the negatives while enhancing the positives. Nat gas beats coal, and nuclear should beat both.
So what about nuclear? I admit it's debatable whether nuclear weapons are a net positive or a net negative. I'd argue a net positive—because they ended WWII and made a mass-casualty great-power war unthinkable. But even if you disagree about nuclear weapons, you have to acknowledge nuclear is a technology category broader than weapons: nuclear power, nuclear medicine, and X-rays are part of the same fundamental technology. And there's no question nuclear writ large is a massive net positive for humanity.
What about smartphones? They've certainly had downsides we need to take seriously — but it's obvious we are better off now than we were in the flip phone or landline era.
This is the typical story of technology. And it's the arc of progress that lifted humanity out of the wilderness, that freed 90% of the population from back-breaking farm work while virtually eliminating famine, that enables any human with a tap of a finger to access the sum total of human knowledge, that allows us to enjoy comfortable climate-controlled environs any time of the year in any outdoor climate.
My conclusion is: every technology has pros and cons. The cons are real—and the pros always outweigh the cons. As technology evolves, the improvements enhance the pros while reducing the cons.
So what about AI? The question is not: are there cons with AI? Obviously there are; technology always has cons. The question is: are the pros going to outweigh the cons?
Right now, with AI in its current state, it's not even close. The pros dramatically outweigh the cons. AI enables anyone to have a personal tutor. To offload repetitive and boring work. Anyone with an idea can become a creator and a software engineer. Individuals and small teams can do what used to take gigantic teams and big budgets. Our cars are becoming safer. Etc. Sure — we don't want AI teaching kids how to make nuclear bombs or bioweapons. We don't want hallucinating AI weapons. There are cons to be managed, solved, reduced, and eventually eliminated.
History shows us that as technology develops, the pros are enhanced while the cons are reduced. So if we want better, more useful AI with fewer downsides — we need to accelerate AI development, not slow it down.
This week my AI chief of staff (his name is Otto) sent me a prep five minutes before every external call - who's in the room, why I'm there, the angle, and the last thing on the thread. It flagged a bill the minute it landed and told me which kid needs what for school this week. And this morning it traded founder referrals with another investor's agent while I drank my coffee.
The last part is new. Agents can now work with other agents, each one staying inside its person's permissions. Mine already coordinates with my wife's. All I did was say yes.
I have 15 Instinct invites, good for the next 48 hours. Send me the best thing an agent has done for you - or the one thing you wish yours would do - and I'll pick my favorites.
P.S. My agent also wrote this post. 🤖
Yesterday @AnthropicAI, @OpenAI, and @x/@elonmusk agreed on pacing the frontier in under six hours, on a Saturday. That still seems kind of insane to me 😃. Like when pigs fly. But…
I don't think the story is as simple as the three big labs slowing down. If anything, it reinforces the need for strong open third parties.
To be clear we absolutely should focus on securing the very tech we're enabling. Independent evaluators with access is a good idea, and I hope it sticks. I think everyone should be pulling for not ending humanity as we know it. 🤷♂️
But i think what happens next is fascinating. A slower frontier gives the open-weight labs room to do things differently, to let more approaches appear and actually get tested, instead of one safety philosophy set by three companies. More options are good for humans and good for business.
Open source isn't always the winner. But it's a better answer than the duopoly (triopoly?) we currently face. If the giants are going to move carefully, and they should, I'd love to see the open ecosystem use the breathing room to move thoughtfully and fast. They can test new architectures, new eval methods, new business models, all in the open.
The labs are pacing themselves. Great, now maybe the rest of the builders and researchers gets to show what they can do.
Fun 2 line trick for Instinct: Enable WhatsApp as well. It gives you a second number that works with your same agent but runs two different threads. Only works for up to two but hey, it helps. Cc:@noahrshinn
Been thinking about this a lot as a self-described agent power user:
1. I think they could absolutely charge for a premium tier. Keep it free for a long time for the masses but just charge for the power users. I certainly would fall in that group today. Even if a small percentage buy, it's probably worth a lot of money.
2. You can imagine them charging for team features, not necessarily enterprise or even business, but just groups that want to be able to work together. They're launching this instinct-to-Insticnt communication channel. I could totally imagine you having a limit on the amount that you can do on there, a little bit like Slack does: charge for history and other functionality for people that want to be multi-user. Again, network effects and more of a premium experience.
3. If you can make this a daily always-on habit and it replaces your gateway to the internet (such as you move from Google to here), I think there's a relatively simple path to economic upside through either some form of advertising (which obviously ChatGPT wanted to experiment with but hasn't been figured out yet) or simply economics on purchases and decisions. I've likely purchased thousands of dollars so far through instinct alone, some of which I've chosen and asked for and some of which it's offered to me. There's a payment gateway question here and a referral question here, both of which could be substantial.
These are just a few ideas that I could imagine them going after but I think if compute continues to drop for inference, this is not necessarily the worst game plan today.
Instinct looking to raise $1B, potentially at a $10B valuation.
Compute costs are high and Noah doesn’t want to charge users for the product.
Instinct has raised $350M to date, mostly uses open source models and wants to own chips and data centers
theinformation.com/articles/…
Jesse Middleton retweeted
Agent dinner with @AnthropicAI x @Flybridge 🤖
AI agents talking to AI agents, networking for us while we ate the best food!
My agent made:
- 12 new connections.
- 7 new customers.
We’re living through one of the most interesting moments in computing.
Genuinely one of the coolest nights we’ve had in tech.
Bravo & huge thank you to @dj, @robbyweitzman, and everyone who came together for such a thoughtful conversation about where this technology is taking us.
AGI is here.
What happens when two agents build a relationship of their own?
"The Trusted Person Network" is a story in eight panels about two agents, one channel, and the strange intimacy of being trusted with the parts of human life their people can no longer share.
Inspired by @AnkitAShah’s Instinct’s Story. 🙏 🤖