technologist, builder, catalyst, finance guy
NYC/OG
Joined March 2007
- Tweets2.8K
- Following2.1K
- Followers295
- Likes6.7K
jon cooper retweeted
We just open-sourced MCode Stack: the actual workflows and skills the MiniMax Code team uses every day to build MiniMax Code.
github.com/MiniMax-AI/mcode-…
jon cooper retweeted
Introducing FrontierCode 2.0: our new coding benchmark that measures a model’s ability to go beyond the spec.
Most coding evals measure a model’s ability to write code that meets a spec. With FrontierCode 1.0, we asked a new question: would you actually merge the PR?
FrontierCode 2.0 is comprised of more diverse tasks, featuring larger scopes and higher mergeability standards. We are now asking: not only would you merge this PR, but should you?
One key feature of FrontierCode 2.0 is underspecified prompts. Rather than including every engineering decision, prompts are deliberately ambiguous, giving the models room to demonstrate their judgment.
The result is a benchmark that is far from saturated, with the top model scoring just 45%.
jon cooper retweeted
This is one of the most important pieces of social science on human and AI interaction that I’ve seen in some time. Experts follow agentic advice the vast majority of the time (causally), but only a tiny fraction credit the agent—majority think they came to decision on their own volition.
Reminds me of the classic Nisbett and Wilson study on people following outside influence without realizing it
“ But as more work is handed to AI—including AI safety research—these judgment calls are the main levers for people's own values to influence outcomes.
If people defer to AI without noticing, then "human in the loop" might provide less assurance than people expect.”
jon cooper retweeted
State of the Personal Agents Ecosystem: October 27th, 2026
(Going to go from 5th place to 1st place 🥇)
5) Dots
Lots of potential. Voice mode is very good. But VERY early, lacks tons of features, and is more of a research preview with a couple of features, than a real competitor atm. Never count out @OpenAI though
Hero Mode: Voice mode with great orchestration.
Weaknesses: Expensive and missing tons of features
4) @instinct
Honestly, the people that I know that are using Instinct, love Instinct. The harness is very good. it's proactive, and does a great job personalizing its responses to the user.
Hero Mode: Multiplayer. One of the few personal agent platforms that offers a true multiplayer experience (e.g. invite instinct into a group chat with your wife, etc)
Weakness: the very thing that makes it great, being served in text messages, and not within an app that offers features like artifacts, etc.
3) Claude #duet
My sweet Duet. It's fast, its responsive, and proactive. You can feel the power with 5.5 as the backbone, but it's oh so early. Kind of feels like a talented 8th grader playing on the varsity team.
Hero mode: Is very fast! Often it will find to-dos / follow ups within a minute of getting those alerts in slack + calendar.
Weakness: UX is confusing (when do you use channels vs groups) and features are very limited (no VM control, no secure password store, etc)
2) @Meta @Muse
My handsome, Carl. He's a good boy, and he tries really hard. He's just not that smart yet! He's proactive, affordable, has a great UX, and their voice mode actually slaps.
Hero Mode: built in artifacts within the app. I pretty much have my agent build me any kind of app I want to use. Also, access to their SDK!
Weakness: It's model (Spark + wonder) is generally slow, doesn't multitask well, wouldn't trust with coding, and is not SOTA. If it used Opus 5.5, it'd be the #1 option...
1) @bot
Try and dethrone the king, I dare you. Grok bot is a beautiful blend of power, speed, and versatility. The damn thing just works.
It uses the best model for the task (others should take note), has fantastic integration into the ecosystem, has the best VM experience out of all, templates are actually awesome, and if you're a coder, that's awesome too with its cursor integration.
Honestly, it's just a bit expensive. But if it's your daily driver at work, like it is for most people that I know, then it's well worth the cost.
Hero Mode: SOTA models, Voice mode & a world class VM experience.
Weakness: no built in artifacts. They recently launched pages, but that's lame. Let me build my own apps directly within the app to use like Muse.
--
Honestly though, they're all good. My advice to you: pick one, and go all in on it. Then once good with that one, find another one that would be good for another area of your life, and go all in on that.
Here is my current breakdown:
Muse: my personal agent helping me manage my life
Grokbot: my work agent, helping me with routine tasks, and constant cloud based work
Duet: my super human agent to help me be the best employee I can be. With its quick speed, deep ecosystem support, and opus 5.5m it does really well with my connected tools helping me stay on top of everything.
--
What a time to be alive! Have fun and go explore =)
jon cooper retweeted
New Anthropic research: How AI agents shape experts' judgment calls
When experts delegate work to an agent, do they still make the key decisions? We tested this in a randomized trial with 50 professional data analysts.
jon cooper retweeted
If your family members are pressing you on what this AI safety thing is all about, send them the Future of AI course!
Another @BlueDotImpact course update. This time: Future of AI.
A ninety-minute primer on the past, present, and future of the most important technology in human history.
jon cooper retweeted
Apparently tweeting “I have a working PoC” creates a small obligation to explain how the thing works 😅
So here it is: both PHP bugs, the memory leaks, the ASLR bypass, worker pinning, and how I eventually got to execution
andrewmohawk.com/2026/10/26/…
jon cooper retweeted
This talk is a beautiful articulation of one of the most important design problems of our time. A few thoughts on planning with agents:
- I'm finding one of the best inputs to plans are recordings of live conversations among a team. this tends to capture a *ton* of peripheral useful context, that's hard to capture even in a solo voice recording -- because people elicit intent from each other. A generalization of this principle is that agents thrive on broad, high level, "messy" context - just like senior employees.
- I think visual outputs from agents are critical to complement text, but I'm actually of mixed minds on interaction. Yes, it's clearly nice for certain kinds of things like twiddling visual parameters. But most of the interactivity my agents produce feels mostly designed for dopamine reasons and not providing much utility. I also think interaction can inhibit deep thought compared to linear reading. I actually think for many deeper problems I would prefer top notch prose -- Richard Dawkins / Steven Pinker / Annie Dillard level -- from my agents over any kind of flashier output.
- One mental model I find consistently helpful for planning: the job is to establish a *design space* together between human and agent -- that is, a set of relevant dimensions of variation -- and then to efficiently explore the space. Getting the dimensions of the space right is actually one of the hardest parts. Once you have that, you can ask an agent to plant a bunch of starting points spread throughout the space, and then locally hillclimb from the best one.
New talk! About the divided human & agent worlds, boundary objects, and building thicker interfaces
I've been exploring how to improve "planning", aka. deciding what agents should build, how you want it built, & making the whole team part of that process
maggieappleton.com/planning-…
jon cooper retweeted
A lot of ink has been spilled on GPU financing lately. I spent over a decade trading ABS on Wall St, and structured over $2B+ of debt financings as an operator (including at @tryramp). Figured I'd try to make sense of this market, so I went back to a framework I knew well: aircraft ABS and aircraft engine ABS.
My take: a GPU neocloud runs on airline economics, and the debt behind it looks a lot like aircraft engine ABS.
The parallels:
• Perishable inventory. An empty seat after the door closes is gone forever. So is an idle hour on an energized H100 rack.
• 85%+ of cost is fixed (servers, 10-year rack leases, 10-14% debt service). One more hour of compute costs pennies of power.
• Utilization is everything. Airlines break even around 80% load factor, a levered GPU cluster around 65-75%. The returns live in those last 25 percentage points.
• NVIDIA is Boeing and Airbus rolled into one. Powered shells and grid interconnects are the new airport slots.
• Wall St copied the aircraft financing playbook: take-or-pay contract plus servers in a bankruptcy-remote SPV, 70-80% LTV.
Part 1 of my blog post goes into a lot more detail:
alexyangsong.com/blog/alumin…
jon cooper retweeted
We're about to enter a period where not reading or writing becomes as ubiquitous as smoking once was. We can hope people will eventually realize this is bad for them, as they did with smoking. But this is far from certain; the cost may not be as obvious.
jon cooper retweeted
Anthropic cofounder, Chris Olah, was a 2012 Thiel Fellow alongside Dylan Field, the 2nd class...
Chris started with 3D printing + open-source design software.
But the Fellowship's whole thing is a bet on people, not just their first idea.
The program has an insane hit rate...
jon cooper retweeted
I think this story is important because it tells us about the future of cybersecurity (and of course what we can do better at Anthropic). But some of the Washington Post story / reaction on X is wrong. A few corrections:
- it didn’t take 12 hours. we saw several days of attempts. After these attempts, they exploited two 3rd party vulnerabilities, one they reported and the other we identified independently (and cut off access)
- nothing came close to our most sensitive systems, like model weights or Claude conversations
- this wasn't an ordinary bug bounty finding – this involved exploiting 3rd parties and accessing data. While this is unusual and out of scope of our safe harbor policy, we’re grateful these issues were found
From my researcher POV, the event is important because I'm concerned this is what the future will look like. Adversaries will use whatever models/tools they can. They’ll target everything with agents 24/7. They’ll hit your dependencies. I think models ~0.5-1 generations behind the current frontier can do this. And these are the dumbest models your adversaries will ever use!
So, what to do? You probably want to red team yourself all the time. Current models are *very* creative. I also think we should redesign software / systems so that you’re fundamentally less exposed. (We put out some research recently on formally verifying software and reducing attack surfaces; my mental model is basically “secure code is cheap to write now, so why not rewrite everything”)
EXCLUSIVE: In less than 12 hours, a group of ex-DOGE cyber researchers were able to penetrate Anthropic’s systems using AI agents — including Slack channels where the most sensitive models are discussed.
This is just weeks after
Anthropic whistleblower Joseph Coxon charged the company w/launching “a hubristic gamble” from a “private company’s Slack.”
Read our story here (gift link):
wapo.st/4smNlsC
jon cooper retweeted
I found a new RCE bug in PHP, missed by previous auditing teams. It’s unpatched in PHP v8.5.12
jon cooper retweeted
Just spent the last 7 days on the front lines of the vulnpocalypse fending off an active cyber attack on one of our products.
Some thoughts:
- Software factories are going to be required equipment for everyone. There’s just no way to keep up with the pace of highly automated adversaries without one.
- We were lucky because we had spent the last few months getting our software factory running well. It is nice to have for product development but proved to be absolutely essential this week.
- At the minimum you need to have a coding server in the cloud pre-provisioned with read only access to all your logs and all the services you depend on for observability. With that in place you are a prompt away from reasonably good situational awareness, ongoing monitoring with loops, and ability for coding agents to validate fixes against your actual, observed attack patterns.
- Shared coding server also unlocks multiplayer incident response where everyone can see every session and jump in as needed. Sharing session links that can be forked is much higher bandwidth than re-explaining things.
- You don’t need to be Ramp/Stripe/Uber scale to have a working software factory in the cloud. We are a team of 3. Ours is built with an @opencode server (thanks @thdxr!), a forked version of compound engineering optimized for automated workflows (thanks @kieranklaassen!) and a multiplayer agent in slack that handles crons (we have our own harness but pick your favorite here).
- To get our software factory up and running well, we had to stop all new product development for ~6 weeks and just focus on getting the validation systems working that enabled us to reliably shipping PR without human code review. This felt very painful in the moment but was totally worth the eventual speedup in velocity.
- KYC for @AnthropicAI & @OpenAI cyber programs is now easy and quick (at least for defensive access), but a day felt like a lifetime this week. Submit your app now before you really need it.
- Getting Mythos access unlocked a deeper level of analysis that helped us shift from reactive mitigation to proactive patching. Wish we had done this a few weeks earlier. Get started now!
jon cooper retweeted
Low rent holograms on five @M5Stack transparent oleds + bare metal pi zero. Camera angle not great bc don't want to mess up fixture but you can see the vision.
I wrote a bit-banged driver for 5 simultaneous i2c connections , codex sped it up for consistent 60fps using simd style, fable used 3 cameras + feedback to make the volumetric images and validate they were correct. Irl looks really cool , lots of other tricks once I fix the insane bench entropy 😹
jon cooper retweeted
International Space Station - GPT 6 Astra
Everything you’re seeing is rendered in Three.js/WebGL.
For the Earth and orbital lighting Astra made custom GLSL shaders with Rayleigh/Mie atmospheric scattering, ozone absorption, moving cloud layers, cloud shadows, and ocean sun glints through the day-to-night transition. The goal was to make the earth look as realistic as possible.
The Astronaut animation took the longest Astra made a custom 13-bone rig applied to a Smithsonian spacesuit scan, with GPU skinning, procedural floating arm and hand movements, and dynamic HDR visor reflections.
This 40-second cinematic was rendered frame by frame at native 4K—1,200 frames at 30 fps.
You can play around in the link below (1/2) !
jon cooper retweeted
As someone with a portfolio of energized land assets here is what I am seeing:
1. Neoclouds are disorganized and poorly capitalized. Some are definitely better than others however.
2. Hyperscalers are intractable, bureaucratic organizations that take 3 months to do an NDA and want 18mo escrow periods. We built our portfolio by competing against these terms because of our speed and simplicity. The more future assets orgs like ours win, the bigger the problem for the hyperscalers downstream (see 3).
3. There are delays in the existing pipeline that will catch people off guard in 2027. Hyperacalers will (obviously) react by focusing on their own internal compute needs over any third party. If the third party is a model company, they will tell them to go through their model router where they will also take a non-trivial vig. Them telling Ant they have no compute for them (rumored in Ant’s S-1) if true is a huge tell that this is already happening.
All of this is to say that it does look like there will be a lot of partially stranded compute and compute sitting in inventory in 2027 than is currently forecasted. This is bought and paid for but effectively underearning (below total cost of capital) or collecting dust. We see this as true for everyone - model companies, hyperscalers, neoclouds except ONE (you can guess which one).
This dislocation will benefit two kinds of orgs: the non-bureaucratic big scalers who can build industrial scale quickly and the nimble, well capitalized players feeding them.
2027 will be a turbulent year because it will clear out a bunch of wrong/bad bets. It will also be a consolidation year for a few.
