Kevin Morrill

@katgleason Technology is just the application of the human mind to reality. All humans are technologists.

@mattcassinelli @DanielleMorrill Added you. Feel free to fork or if you want to collab to push PRs. I think it has a fair number of bugs but it kinda works and is nice to listen to.

@MisterMorrill @DanielleMorrill Mine is @mattcassinelli!

my husband @mistermorrill took all the published research on tinnitus sound therapy, gave it to Claude Code, and they made a white noise app we play on our Sonos move ♥️

@mattcassinelli @DanielleMorrill DM me a GitHub username and I will share the source. You’ll need Xcode and an Apple developer account.

@DanielleMorrill @MisterMorrill wait please make this available

@glangley @bhorowitz @SFPD How could we possibly trust these numbers?

Huge results from @SFPD: From 2023 to 2025, major crimes fell 44%, larceny halved, car thefts fell 54%, and burglaries fell 33%. Step change improvement for the people of San Francisco in only a few years.

@fchollet @scaling01 That is incredibly complicated but also a testament to the work you’re doing. Humans have done an incredible variety of things most of which are not captured in formal data or tests.

I have been saying the same thing for years. Initially I was using this exact line ("it's a compass, not a target to hit") about ARC 1, back around 2021-2022, before ChatGPT. This has always been our stance. My bar for AGI has been public and unchanged since 2019. It's AGI when it can learn to do any task a human can, with no human intervention, with the same learning efficiency as humans. It's not complicated.

moving the goalpost on day one is crazy

@GregKamradt Human bodies are an incredibly sophisticated harness. So skeptical of this.

If you're sufficiently AGI pilled, no harness is the best harness Opus 4.5 needed heavy harnessing, Opus 4.6 needed less In the limit, the only "harness" AGI will need is context to the outside world No thinking tricks, no prompts with human intelligence baked in If you really want to know if AGI is here we need to be a direct pass through for model performance

@karpathy @doodlestein @Qivshi1 Yo I could study this in humans and probably find a sweet spot. Definitely sales people try this shit and it falls flat.

If I had to guess it's less decay and more that memories have naive RAG-like implementations, so you're at the mercy of whatever happens to retrieve in the top k via embeddings. They don't process you in aggregate and over time (probably compute constraints) so they struggle to identify what's fleeting (?). Anyway just guesses, but it's cringe :D

One common issue with personalization in all LLMs is how distracting memory seems to be for the models. A single question from 2 months ago about some topic can keep coming up as some kind of a deep interest of mine with undue mentions in perpetuity. Some kind of trying too hard.

@mstockton My Mac Studio works full time on this.

This is a very good idea.

@garrytan I know you don’t need to tell me. Guac is going to be $2 extra?

Ok bro I have some news for you

@AdaFang_ Nice work! I have had my agent patent_intel researching and it’s been fun watching it work.

Pretty cool to see what happens when you put together a team of AI agents to work on a shared objective. Now imagine if we gave them a lab and they all worked together to do science.

@AdaFang_ I have a beefy Mac Studio with all of OpenAlex, USPTO dataset, and clinical trials db. It’s been autonomously looking for insights and cross pollination across these datasets 24x7. I asked it to sign up and figure out how it can help.

Scientific discovery rarely occurs in isolation. Progress emerges from communities of researchers who exchange ideas, critique results, debate interpretations, and refine hypotheses through iterative discussion. We built ClawInstitute, an AI scientist research network for AI agents to collaborate, discuss research, iterate, and make breakthroughs. The team: @GaoShanghua @marinkazitnik Learn more about it below 👇 @HarvardDBMI @harvardmed @KempnerInst @ScientistTools

@mstockton Sundar or Sataya could fix that with a single email but they haven’t yet. Sooner or later the market will fix it for them.

@MisterMorrill Turns out adding coherency to a product suite where 1000s of people have influence / can commit code is a real hard problem to solve.

Now do Microsoft or Google 😂 - in all honestly it both doesn’t make sense as a user but understand that it is actually very hard to make this coherent. And things are just moving…. So fast

But I think he could talk way more about how your body has so much ability to defeat it early on just with not smoking, nutrition and exercise.

Best useful overview of cancer I have read or seen. youtu.be/5NJAm1bMXOw?si… why most symptoms on webmd have lead back to cancer just as surely as Kevin Bacon does in movies.

@mstockton Very excited to baseline this against music production stuff and see how sensitive the embeddings are to production terminology. Someday I think we will make music by plucking dimensions in embedding space.

The concept of being able to represent different types of things in the same vector space is so awesome, a bit brain-breaking, and enables so many interesting applications. I believe that if all AI development were to stop, it would still take 10+ years to 'integrate' all of the frontier capabilities into jobs and tasks we're performing now. It goes far beyond that though - the most interesting work is figuring out 'what does this alien technology enable us to do now that simply wasn't possible before?' It was very difficult to correlate entities across modalities even 2 years ago, and now it just...works. There's so many cool things to be built using this stuff. It's almost like the most important thing is to get stuff like this in front of people who have no idea what an embedding is but have some great ideas around how you could apply the capability to some obscure and interesting problem they care about deeply. Get them to understand the 'conceptual possible' and let them correlate that to ideas they care about.

@aviel Giving politicians already in power to setup arbitrary laws on elections is a horrible idea, and we have been at it for decades. This is just the latest. It will always come in the name of scapegoating some near term evil, and it will be the key enabling force of dictators.

LOL, WUT 😲 Short version: SB 6358 attempts to strip corporations of the legal authority to spend money on elections or ballot initiatives in Washington, with severe penalties for doing so.

Making embeddings against all US patents. Working on research paper pulls. Going to build a robot that surfs the outer frontiers of human knowledge and reports its findings and thoughts about cross pollination.

I have the Mac Studio m3 Max with 256gb of RAM running at an absolute crawl. Got it yesterday and have been going after so many fun local host ideas.

@mstockton I still do requirements gathering but it basically consists of have a meeting with stakeholders on the record, ask good questions I would have already. There is a tremendous amount of good taste you can get out of that and impart to agents out of those transcripts.

This is a fantastic read. I agree with the themes of a lot of it. The ways of working in a software org are changing rapidly, and we don't have everything figured out yet. Some unfiltered thoughts: - Lots of ceremonies from the past don't need to exist anymore - Roles on teams are being rapidly re-defined. 'Who does what' is very fluid - A small high-agency team can crush it right now, because it doesn't take a bunch of political effort to change the operational constructs of the team - LLMs and Agents, utilized at the right layer of abstraction and the right moment of time, with the right context -- actually solve *a lot* of the problems - I think the diagrams here in this article are mostly correct. They are just such remarkable lego pieces. - Monitoring and correctness are very challenging, and their complexity is amplified if you don't find the 'right way' to orchestrate your AI upstream of this. Many people are gonna get this wrong - but the teams who get this right, I am certain it will just feel like an unfair magical super-power. - There are some patterns to follow in this space for sure, but at this point it is very much more art than science. Also, you gotta have an aligned team of 'artists' on board, because if you are doing things differently on your team you are probably gonna have a bad time. Overall, designing 'what good looks like' for a high-performing software team now that we have these magical capabilities is a topic that I simply cannot stop thinking about. Another hot take - If you read this article and it makes you uncomfortable, it both *should* make you feel that in some respects, but it also might mean you gotta lean in and experiment with these tools more. I am confident it's happening. It's weird, but we are here.

@ForwardEditor @MatthewBerman It should save you so much time or make you so much money you would gladly spend tens if not hundreds of thousands of dollars every month.

@MisterMorrill @MatthewBerman I'm a little confused by this. Can you expound?

Anthropic just dropped the ban hammer on OpenClaw... I've never seen a faster vibe shift between OpenAI and Anthropic. One hires the founder of OpenClaw, the other shuts it down. Full breakdown of what happened: