Talks about LLMs, Cloud & Data | Views are my own only if they are interesting, otherwise they're someone else's
Paris
Joined October 2023
- Tweets123
- Following180
- Followers42
- Likes311
Arthur Verrez retweeted
1/5 TLDR; We used Codex to discover and maintain heuristic learning for hard fluid dynamics control cases.
I’ve been applying DRL and GNN to physics since 2019,, and over the past 3 months I’ve been toying with the idea of using agents in our processes. Inspired by the blog post from @Trinkle23897, I decided to use the same strategy and have agents find readable control strategies. This means a lot to our field, where interpretability can be key for industry.
Arthur Verrez retweeted
very small week end side project, forked from autoresearch @karpathy : same idea but using multi armed bandit scores to optimize multiple objectives at the same time
repo: github.com/DonsetPG/autorese…
I don't care about your insta discover page, show me the last wikipedia articles you've checked and let's see what it says about you
Technical tests or case studies that can be one-shotted by o3 are already obsolete.
Candidate value isn't raw capability anymore, it's their ability to surpass what's achievable by top-tier LLMs that matters.
Being able to run locally on a 5 year old laptop (GTX 1650 Ti) a voice model that clearly outperforms OpenAI's TTS is insane (even if it took 4 min and 30s for a 10s generation)
Big up to @_doyeob_ and the Nari team
Your agent infrastructure isn't good enough to solve your problems?
No worries, I got an easy fix for you, just make sure no one's looking
Ok it's rant time, this might make some people angry, but no, solving your current software issue probably does not need LLMs and even less probably an Agent.
I see way too many devs treating “add a manager‑agent” as the default instead of checking the algorithmic possibilities. Classic data structures + two IFs are cheap, fast, predictable, and they don’t melt your openai bill
If your job is plain classification and you own a labeled dataset, logistic regression or a good random forest will out‑perform the token furnace on cost, latency, and uptime. Ship that, sleep at night
Reserve LLMs for true unstructured chaos: parsing messy PDFs, summarizing chat logs, stitching multimodal junk into knowledge
Replacing deterministic workflows with “manager agents” that decide which tool to call next makes me so angry because it's SO inefficient and costly
Search, ads, fraud detection, and routing hit planet‑scale long before transformers. Respect algorithms first: deploy LLMs only where they’re the only thing left
Every 10 to 20 messages, Gemini 2.5 pro exp needs a bit of pushing to get things done on @cursor_ai...
What's stopping you from dropping all your WhatsApp history with a friend in @NotebookLM and getting a podcast of your relationship?
It's terrifyingly easy if you guys want to try it:
github.com/ArthurVerrez/what…
OpenAI's codex actually seems pretty sick, even though I'm a bit scared to give rm -rf access to some LLM that might be gaslighted by some GitHub repos though
Arthur Verrez retweeted
Sorry, but your multi-agent system is still just prompts, loops, and Python functions! 🙃
Btw, thanks @replicate & @bfl_ml for providing the best infra & models for fine-tuned text to image services
Decided to open source this also: github.com/ArthurVerrez/repl…
After testing CrewAI & Microsoft AutoGen, I can say that Google's ADK (that dropped 5 days ago) is the most straightforward to put in place
