Miguel Fernández retweeted
Brilliant paper from Google on RSI.
(bookmark it)
Recursive self-improvement (RSI) lets an optimizer agent improve an agent harness on its own. It reads failure traces, edits the prompts, tools and completion checks, and keeps whatever scores better on a benchmark.
This works until the model underneath changes. A harness evolved on Claude Opus, moved to Gemini, scored below Gemini's plain starting harness in every domain Google tested.
When the harness moves to a new model, the only thing carried over is the evolved harness code, meaning the prompts, checks and rules the optimizer added. The record of how each piece got there stays behind, including which failure it targeted and whether it fixed that failure when tested.
So the optimizer working on the new model inherits a set of changes with no reasons attached. It ends up retrying experiments that already failed, or removing fixes it doesn’t understand.
Google's fix is Self-Improvement Memory (SIM), a log kept for the optimizer itself. Each entry is a recurring problem, like an agent declaring a task done before its commands ran, and every fix attempted under it stores:
1. The edit and the reasoning behind it
2. The tasks it was predicted to fix, written down before testing
3. The tasks it actually fixed, failures included
4. The model and benchmark it was tested on
The whole log goes into the optimizer's prompt every round, and new attempts get appended as optimization continues.
The paper shows the difference on one inherited check that stops the agent from finishing early. With SIM, the optimizer read why the check existed and extended it.
Without SIM, it rewrote the same check four times, twice to fix deadlocks it had created itself.
On the Opus-to-Gemini workspace tasks, the run with SIM improved about 3x more from the same starting score. It also finished ahead in 9 of 10 independent coding runs.
The paper is at arxiv.org/abs/2611.18856
I also wrote an article on a related idea, covering how to make an agent harness diagnose its own failures.
It's built on @Cometml's Opik, which is 100% open source. It goes beyond reading trace logs to finding why a failure happened and stopping it from recurring.
That diagnosis step produces the kind of root causes SIM stores.
The article is quoted below.
Miguel Fernández retweeted
An engineer on our team took a brand-new open-source product from 7.7% to 50.3% of AI answers in its niche. In 3 months.
#1, ahead of OctoPrint, Mainsail, Fluidd and Manyfold.
His part: 5 to 10 minutes a day approving. Omnia Agent did the rest.
5 lessons 🧵
Miguel Fernández retweeted
The man, the myth, the legend @vmg does it again:
cursor.com/blog/git-at-any-s…
Miguel Fernández retweeted
¿Te imaginas que tus hijos se quedaran sin cole de un año para otro? 😔
Hoy no tuvimos patrocinador, así que, aprovechamos la oportunidad para dar voz a las 200 @FamiliasOsotu, que llevan 18 días en huelga de hambre.
Ojalá @BegonaPedrosa escuche esto: youtube.com/watch?v=k9Nx7sql… – at A Coruña, Spain
Miguel Fernández retweeted
uh oh. massive supabase outage. @supabase
Miguel Fernández retweeted
Chinese ByteDance just announced OmniHuman.
This AI can make a single image talk, sing, and rap expressively with gestures from audio or video input.
10 wild examples:
1.
Miguel Fernández retweeted
I got early access to ChatGPT Operator.
It's OpenAI's new AI agent that autonomously takes action across the web on your behalf.
The 9 most impressive use cases I’ve tried (videos sped up):
1. Ordering dinner ingredients based on a picture and a recipe
Miguel Fernández retweeted
Nice and concise read about distributed systems 💯
So much of this resonated with me!
somethingsimilar.com/2013/01…
Miguel Fernández retweeted
This 600-year-old painting is one of the most mysterious in history.
That mirror at the back is just 3 inches wide — yet it reflects the entire room in immense detail.
Look closer at it and you'll realize nothing is as it seems… (thread) 🧵
Miguel Fernández retweeted
35GB is a joke for any database. There is no need to change anything.
Always find out if something is really slow or you just think it is. And when you found slow queries, just optimize them.
Suggesting to switch the database technology is the worst feedback you can give.
Miguel Fernández retweeted
Open source LLMs reaching GPT-4 levels way earlier than we thought is the most exciting thing now to me
With open source RAGs (plugins for LLMs) we can even add stuff like GPT-4 Vision, web browsing and data analysis and reach features parity with GPT-4 and future versions
This democratizes the entire LLM world and means it's possible it won't be monopolized by 3 different BigTech giants but be usable for everyone without any filters and restrictions
This means so many more creative uses of it without limits set by tech companies in San Francisco
It'll be a lot like what Stable Diffusion did for AI image generation
We just need to make open source LLMs easy to use now for normal people
LMstudio dot ai is a nice start but still too complicated to use, eg you have to search and download and pick your own model weights. Just preset it for people! Too much open source engineer thinking and not enough normal user thinking IMHO
Anyway very exciting time
If u want to try, get LMstudio and download the Mistral 7B as well as the new big Mistral model and try it out, free and open source and NO restrictions
Miguel Fernández retweeted
A sync is still a meeting. You've just found a more palatable word for wasting time.
Miguel Fernández retweeted
🎉 Exciting day at Prisma! Today, Prisma Accelerate is officially in GA 🥳
🌍 Its managed connection pool is perfect for Prisma ORM in serverless and edge functions. And its edge caching layer takes your data global, drastically improving latency. pris.ly/blog/accelerate-ga
Miguel Fernández retweeted
The next big recruiting platform in tech is the 32 hour work week 🙏 few understand this.
Miguel Fernández retweeted
Prisma provides a server-side library that helps developers read and write data to the database in an intuitive, efficient and safe way. Here's the story on how they were able to make massive distribution cost saving, using Cloudflare R2. And you can too! cfl.re/3ZPSuLN
Miguel Fernández retweeted
Building data-driven apps has become increasingly complex. I and the entire @prisma team are proud to share an approach that's long been at the heart of our work before it had a name: Data DX.
🚀 Excited to unveil the #DataDX Manifesto: a bold new path in data-driven app development.
More than words, it's our call to action, empowering devs and inspiring next-gen product innovation.
🌐 Dive into Data DX with us at datadx.io
pris.ly/blog/datadx-manifest…





























