Dimabytes

A real problem we have at our YC startup right now: How do we use our AI development budget in the most efficient way?

Opus 5 just disappointed me today. I asked it to build a per-minute win-prob dataset from real Dota2 matches. Opus decided: Radiant kills = all Dire deaths. And forgot about neutrals, deny, suicide... Looks like we are still far from "don't read the code, man"

People are still struggling with effort vs model I recently bumped into one of the best short-formed explanation for this from Anthropic. But it applies to any model, obviously If you want an agent to work more thoroughly and carefully through a problem: Raise the effort If the problem is especially difficult and you want to give an agent more capabilities: Change the model

I made a fork of the cursor-thermos plugin for Codex. It runs two agents in parallel for correctness/security and code-quality reviews and synthesizes one report. This plugin is the best thing I've found in months. I use it for almost EVERY feature I do. But I was missing it in Codex. Now I have it, and you can use it too! Do planning with Opus 5, implementation with Grok 4.5, and review with Sol. It's a proven tactic to find most of the issues. Greptile recently published research about it: "Models are worse at reviewing their own code". Try it: codex plugin marketplace add Dimabytes/codex-thermos --ref main codex plugin add thermos@thermos

Just watched the new Spider-Man movie. And he is actually a vibe coder lol: - Talks to AI by voice - Doesn't really check what the AI is doing (it made a chip and he just put it in his head lol) - Has 10+ terminals open at the same time

Grok 4.5 is great. But why is it ignoring the tool to ask me questions? Why is it always asking me to answer in a follow-up message with a format like 1. B 2. C Makes me think it has some problem with tool use...

Will Codex ever return the 5h limit? I like it without the limit. Cursor has done it since the beginning. Let's see for how long Claude will have it

Switched our org today from Qodo to Cursor Bugbot The weird thing is that every developer needs to set the review trigger type in their personal settings. I can't do it at the org level. And by default, the trigger is "Every push". BugBot charges around $1 per review, btw.

I had to ban Claude from Slack today. Why: We are migrating to GitHub, and shared a migration guide in the channel. Claude was never tagged/mentioned. But it put 👀 on the message with commands. I made a screenshot of that, with the comment "what hahaha". And it triggered Claude to message "On it — checking the local clones and switching the origin remotes". This was super weird, so I stopped it and disconnected Claude from Slack. @bcherny @claudeai fix it pls

I just burned 60M opus tokens in Cursor completely for free Tokens you spend on cloud agent env setup is are free until 5 corrections (and then you can just start a new chat and continue setup) Everyone is talking about Codex resets, but Cursor just cooking silently lol

AutoGPT (160k stars) author is now in YC! Silen Naihin is in YC’s Summer 2026 batch with Experiential Labs. Here’s what they do, in plain terms: If you’re building an AI agent, you need to test it thousands of times. And every test run means spinning up real infrastructure: servers, browsers, databases, APIs. It’s slow and it’s expensive. Experiential Labs replaces all of that with an LLM that simulates your infrastructure. The agent thinks it’s running commands on a real machine, but the responses are generated by a model trained on recordings of how your real system behaves. The result is 5x faster than real infrastructure, and the simulated responses match the real ones 99.7% of the time. If anyone was going to build this, it makes sense that it’s him. He was on the founding team of AutoGPT. The open-source project that kicked off the AI agent era, one of the fastest-growing repos in GitHub history. There he created the first open benchmark for testing agents, turned it into a NeurIPS paper, and got quoted in the NY Times and Wired for that work. Then a YC company of his own (Stackwise, W24). Then AI for science at the Department of Energy. So this isn’t a pivot into a hot space. “How do you actually test an agent?” is the question he’s been working on since 2023. He is just answering it at a deeper level now. His co-founder fills in the other half: Kion Fallah, PhD in ML, previously Staff Research Scientist at Waabi, where he led the program for testing self-driving trucks in simulation. Self-driving hit this wall years ago. Testing in the real world is too slow and too dangerous, so the whole industry moved to simulators. Now the two of them are applying that same approach to AI agents. And notice their first move: they open-sourced the core of the product (world-model-harness on GitHub). The same playbook that started AutoGPT. If you want a way into AI, this is it. Not a perfect CV, but public code. Open source is still the best career move in this field.

I’ve been having sleep issues for the last few months. I’ve been trying to fix it by adjusting the temperature in my bedroom, but at this point I don’t think there’s much more I can do. My AC is turned on at maximum fan speed and set to 16°C. When I turn it on, the humidity drops below 40%, so I bought a humidifier and run it at night. But it’s still not helping. I go to bed around 23:30 and usually fall asleep around 00:00. Then I wake up multiple times during the night. Like at 3am, 5am, and 7am. I usually fall back asleep, but it’s in this weird state where it doesn’t feel like deep sleep. I also get thirsty every time I wake up, so I keep a bottle of water next to my bed every night. Another thing is that at the beginning of summer I was sleeping under a thin blanket, but when the AC is on and the temperature goes below 22°C, it feels too cold under a thin blanket. So now I sleep with the same blanket I use in winter. But it can be really hot. At this point I’m thinking about getting my AC checked to see if there’s anything I can do to improve its performance. It’s a bit difficult because the building is from the 1980s, and I already have PVC windows. My AC is about three years old, and I’ve never had it serviced. Another thing I want to try is a water-cooled blanket, but it costs like one Fable session (a lot lol). I don’t know if anyone has dealt with something similar and can give me some advice. I’d really appreciate it

I will do some interviews with BE devs next week. I changed the approach for the live coding part. Instead of LeetCode-style tasks, I will prepare a real-world task related to our field and structure it like this: - 10 minutes: intro with the candidate about the task - 1 hour: the candidate builds it using any tools available - 10 minutes: discussion at the end - 10 minutes: their questions Still not ideal, but I think it's 100x better than banning AI tools in a coding interview. The day-to-day work of developers is changing, so the interview process should change as well.

If a company allows the use of AI tools during an interview or a take-home assignment, who should pay for the tokens used? I think the candidate should pay for the tokens. Because the company doesn't cover the internet, computer, etc. Why should tokens be covered?

Ernest Hemingway was challenged to write a tragic story using only 3 photos. He won the bet with: – at Amsterdam, The Netherlands

Just skip today...