sam

My FireRed bot can now do the first 2 gyms and it generates this cool gif from each run

Claude terminal UX for agent/session mgmt is pretty bad, why is it even possible to accidentally resume the wrong session and immediately tank a 5hr token limit?

can someone tell me all the things wrong with doing `echo "$PASSWORD" | sudo -S apt install -y some-package`

Love it when Opus gets a little poetic: "killing this run, it's doomed"

A Claude Code pattern I'm enjoying when we're waiting for a long run to finish in the background, ask for check-ins every 30 minutes

Great paper

used a ralph loop so that my pokemon bot can complete the Nugget Bridge farm as fast as possible (could still be optimised a lot tbf) blog.samgould.net/post/agentic-o…

nice one thanks Opus

@benstein @simonw 1-3 feel very Byung-Chul Han

Hey @simonw we need you to coin some new terms! Here's 4 new phenomena I've seen in 2026 that we don't have words for yet: 1. The constant feeling of anxiety / FOMO / guilt when your agents aren't running. Just one more prompt! (my working name is Token Processing Underutilization Disorder, or TPU Disorder) 2. The desire that every founder/builder has right now to solopreneur a company with just Claude 3. The desire that every founder/builder has right now to solopreneur 10 products/companies at the same time 4. Our newfound ability to fix/improve every. single. thing. you think is broken in the world. (This month I made a personalized weather app for my watch, a replacement app to control my heat pump, and a google calendar of the Oakland Ballers home game schedule. Just because. I can't stop.) Whatcha got??

@sethkarten Such a cool paper. If I understand correctly it can write arbitrary Skills? Could the same approach be used in a constrained way e.g. to optimise choices of chunking and retrieval strategies in a RAG agent?

Improving your SKILL.md is not a stationary optimization problem. Continual Harness reveals inherent issues that occur when self-improving skills. 1. Skills can oscillate in performance quality especially when new data reveals a covariate shift. 2. And even when skills strongly dominate basic inputs (think bash/button interface), the model refuses to use them over time The only way to improve this is to co-train the harness and model to make a true foundation agent

AI has now solved a major open problem -- one of the best known Erdos problems called the unit distance problem, one of Erdos's favourite questions and one that many mathematicians had tried. openai.com/index/model-di…

But the tl;dr version is that the model proved a result that in my assessment would have made a perfectly reasonable chapter in a PhD thesis. It did this in a total of a couple of hours, with a few prompts from me that contained no mathematical input whatsoever.

Last month I wrote about how I'm using AI agents, and especially how I've set up my harness for agentic software engineering. This month I'm focusing on how to use these tools to design software that can complete objectives - favouring the abstraction of scriptable macros rather than a fully autonomous LLM brain - and applied to a certain video game. blog.samgould.net/post/building-…

@ianlapham What do you use it for?

After 2 months of everyday use, I can say that setting up a personal research engine is one of the highest-ROI things you can do if you like to learn and stay on top of things at the edge - Use a cloud-hosted agent, probably hermes or openclaw - Learn about memory systems and encoding (cognee is very good at this) - Build the right commands for parsing data and storing it (tag things properly, encode and save full text and key ideas) - Build recurring jobs so the system grows itself (rss ingestion, auto twitter scroll, newsletter following) - Build advanced skills that create connections between ideas, surface the most important info, and create digests for you automatically - Build search retrieval skills that actually pull what you need and don't forget or miss things Will change your life

@bee_bebetter I configured mine like isaak.net/mandarinmethod…

@bee_bebetter You shouldn't have 100s of unknown words being shown to you, have you configured to only see ~5-10 new words per day?

Anki doesn’t work for me. I’ve tried using it many times, but it just doesn’t work 😭 the pressure of having hundreds of words on a deck I haven’t memorized, is overwhelming. I might get them right after repeating them more than twice, but after that, I forget them😭

There's a big difference between how programmers and non-programmers use agents. The former default to using agents to build apps/scripts to complete tasks in a repeatable way, while the latter want to chat while the agent completes the task in realtime.

THIS GUY GOT TIRED OF MANAGING AI AGENTS THROUGH TERMINALS AND DASHBOARDS SO HE BUILT THEM AN RPG WORLD 5 agents and each one has a pixel character, a station, and they actually walk around the space when enough unresolved issues pile up, the agents walk to a meeting point and hold a council session. four different models debating what to do next, not scripted. each one reads the live system state independently. in one session an agent pushed for cold outreach to close leads at 2am. another one said that's a terrible look for an autonomous system contacting strangers while the operator sleeps. they ended up pivoting to an inbound strategy that none of them originally proposed. single HTML file, node bridge, and phaser. runs on a Mac Mini. instead of reading logs and checking dashboards you just watch your little pixel agents walk around and talk to each other this is the most creative way i've seen anyone manage AI agents so far

If you add a Claude hook `powershell.exe -Command (New-Object -ComObject SAPI.SpVoice).Speak('Task complete.')` then it speaks to you every time it finishes a task. Could even pipe in the message content.