Agentic Development - https://teapawt.org/t.co/Z2OHCqrz8N
Ireland
Joined June 2007
- Tweets1.6K
- Following581
- Followers1.6K
- Likes708
Did @OpenAI Astra just reverse engineer another AI, what the heck is Weight Space Programming and why is everyone talking about it and why this might be a big risk
Congrats to @rikiyatakehi , @jayden_teoh_and , @a1zhangthe and @lateinteractionteam on this novel approach and detailed balanced paper and to @NeelNanda5 @giffmana @mustafaergisi for some of the original ideas supporting the breakdown research
Huge if true, hopefully it gets replicated
We asked GPT-6 Astra to edit Qwen3.5-0.8B weights to improve math.
It directly changed only 0.016% of weights, fixing 0.8B's reasoning loops & improving math up to 6x and even instruction following 3x, without gradients or training.
We call this Weight-Space Programming: frontier LLMs can now natively read/write neural network weights! 🧵
Blog: rikiyat.github.io/blog/weigh…
Killing the coding interview feels right. I'm seeing this more and more.
It's just not useful anymore. The next question is what skills do you interview for, and how?
teapawt.org/felixrieseberg/status/…
The coding interview is dead, at least for my team at Anthropic. I've seen enough to know that it no longer provides signal. I removed it from loops this month.
I have a spicy, optimistic take - I think there's an opportunity for junior engineers: felixrieseberg.com/hunger-be…
It's pretty clear now that good prompting, and now good software factories, are simply good management principles in disguise.
Clear objectives. Don't micromanage. Provide the knowledge and tools required, then stand back. Don't step in to fix it; empower the agent (or employee) to fix it. Lock mistakes and learnings into SOPs (Skills).
Lead, but get out of the way as much as possible.
Opus 5.5 is such a beast. It's dominating my usage the last couple of weeks. I'm hardly touching other models for my day to day. Nothing compares to it's proactivity and follow through.
A good harness is harder than you think. It's the combination of many small decisions and improvements over time.
For example, in @cursor_ai and @bot, instead of being blocked by the next step in a typical IO fashion, the agent makes good use of its time and parallelizes approaches.
Grok Bot in particular is good at this. Instead of starting a new thread, you can keep chatting in the same one while it pushes the previous process to the background.
Most useful benchmark advice I've seen this month.
Benchmarks tell their own truth. What matters is how the model works for your own use case.
@warpdotdev seems to have made some big savings by picking the right models for the right task. Worth a read.
teapawt.org/zachlloydtweets/status…
I loved this man and what he represented. A fantastic progressive future. An eloquent and educated role model. I’m going to pour an earl grey and watch “the inner light” this evening in remembrance.
Google sharing benchmarks but not the model is like talking about your hot Canadian girlfriend no one has ever met
No shade towards my friends at Google, it's not their call, but I just think these model teases are seriously one of the most annoying things you can do.
No one wants to hear about a model they can't use.
You'd be better off just going to your trusted partners directly with access, and allowing the leaks to generate hype for you.
A model I can't use is a model that I do not care at all about, because it doesn't exist.
It’s been 2 weeks since I’ve used an @OpenAI model for any software dev. ChatGPT was my go to for quick questions and that’s been replaced by @bot
They have great models but @AnthropicAI is just consistent and better and Grokbot is such a great interface for project work and adhoc questions.
This looks really interesting if true. Not sure about some of these claims, but worth a read if you experiment with agent workflows.
teapawt.org/jsnnsa/status/21156240…
i've been building the most ai-pilled company in the world for 2 years. let's trade notes:
- spawn is ten people
- we spend 30x payroll on AI tokens
- and ship almost as much as anthropic
- our agents pick their own names
- and profile pics
- and remember everything
- they dream
- and have time off on sundays to relax
- (not joking)
- agents fix every error hit by a user
- and verify their own work
- with no human in the loop
- we don't care if they break production
- only that they fix it fast
- we let bugs design how we work
- not code bugs, ants
- only anthropic's models can do the job
- openai makes reward hackers
- xai and google make village idiots
- i wrote an article with all the sauce
- but you'll scroll tiktok instead of reading
I have to say Claude Desktop has the nicest little design surprises of all the agents.
It's the subtle animations. They feel considered.
This is how we are going to bring Stack Overflow back. We start using it as a task manager for our agents. Every Agent posts a question and the subagents have to respond via overflow comments. #bringbacktheflow
I covered Superpowers by @obra over a year ago. It's gone from strength to strength, now 300K+ stars on GitHub.
If your team follows a traditional SDLC, it's still one of the strongest spec-driven approaches out there. It came out on top in my YouTube review series.
Still going strong. Check it out 👇
github.com/obra/superpowers
It's hard to decide if I move faster in my business by compounding my agents or hiring another AI pilled operator like myself to assist in the build.
