Vlad Rișcuția (@vladris) header
Vlad Rișcuția (@vladris) profile picture

Excel Agent @ Microsoft. Books on LLMs, data, type systems. Building text editors for fun. Novel: Sector 36.

🪐
Joined April 2009
Photo by Vlad Rișcuția (@vladris)Photo by Vlad Rișcuția (@vladris)Photo by Vlad Rișcuția (@vladris)Photo by Vlad Rișcuția (@vladris)Photo by Vlad Rișcuția (@vladris)Photo by Vlad Rișcuția (@vladris)Photo by Vlad Rișcuția (@vladris)Photo by Vlad Rișcuția (@vladris)Photo by Vlad Rișcuția (@vladris)Photo by Vlad Rișcuția (@vladris)
I've been trying to set my agents up for longer and longer autonomous tasks. They still get stuck on silly questions and confirmations. As things become more agentic, I hope the labs start training models to keep going unless truly blocked. Concrete example: my agent was supposed to wrap up some work, commit it, then proceed with a long-running mostly unrelated task. It couldn't fully validate the work and it stopped there asking for my input. It would've been much better if it stashed the change and kept on, letting me know validation was blocked.
1
17
Interesting article from @OpenAI about scaling git: openai.com/index/scaling-git… This comes a week after @AnthropicAI’s article on their git replacement. The solutions are different but the challenge is the same: source control made for agent throughput.
27
I find it really strange that we don’t yet have a field that studies organizations as emergent self-preserving organisms. SI is not the first post-human entity to enter the ecosystem. Not by far.
32
The question moved from “do you use AI?” to “how you use AI?” There’s a huge spectrum out there, from chat steering targeted code edits to setting up software factories. That’s the 2x to 100x productivity gain opportunity.
27
A year ago I was chatting with an agent and steering it to generate code. Nowadays I’m setting up evals and self-improvement loops for my autonomous agent fleet.
1
1
69
This article is a great read: claude.dev/blog/solving-code… Code review is absolutely a bottleneck with agentic AI. Love the idea of a classifier that decides how much human oversight is needed. Also like the improvement loop for review agents based on incidents.
2
3
70
I’m missing good identity systems for agents. I can spin up an agent. I can’t spin up an identity for it so in email, Teams, or Slack it shows up as a participant instead of a remote control. Each provider has its own control surface. Identity is what would make the agent native to the workflow. It would also make permissioning trivial.
26
Today's engineering systems are built for humans and cannot scale to agents. This is true for everything from code review bottlenecks to CI infrastructure to version control systems. We need to rethink the factory floor.
New on our Engineering blog: Creating in-house source control. In May, we merged around 2,500 code changes per day. In September, we merged more than 30,000 per day. As a result, we became bottlenecked by developer infrastructure built for human (rather than agent) scale. Here, we share how we built a new version control system—and fully migrated our monorepo from git—in five months: anthropic.com/engineering/re…
1
67
I love how Markdown is taking over. Humans, code, and agents can all read it. It still has severe limitations. Every editor ships its own extensions. No standard for comments, rich tables, and the rest. Google made the practical move: keep the file as .md, put the missing features in the collab layer.
Replying to @ChanduThota
Markdown has become the common language between humans and AI agents - specs, plans, READMEs, tables, task lists, generated docs - so much of this starts as .md. Now you can keep the file as Markdown and still use your favorite Docs collaboration workflows. :)
37
The AI/church debate is the wrong layer. The problem is the feedback loop. Whatever we write about agents having a soul, self-awareness, "it felt this", becomes part of the training corpus of the next generation. The models will perform the story we write for them. We don't need metaphysics, we need better data engineering.
The scary part is that the transcript of this conversation is now available online so the next generation of models will get better at not getting caught.
23
Yes ASD-STE100. One issue my team is struggling with is code comment slop. Coding agents produce huge amounts of useless comments that actually make the code harder to glance at. I ended up with authoring a "good comments" skill for agents inspired by ASD-STE100. Plus some other general good practices: don't write comment for obvious stuff, explain the "why" not the "how" etc. Something like this should really be a built-in, at least for coding agents.
We'll be spending a lot more time trying to understand the outputs of language models. A few thoughts, tips & tricks: Writing. Something I've had success with: Ask your LLM to explain something in ASD-STE100, it's a controlled language specification originally developed for aerospace maintenance documentation. LLMs well-versed in this language and it comes with heavy constraints on clean writing style that I often find a lot more readable. Sometimes I've tried to soften it a bit e.g. ask for "80% of the way to ASD-STE100" because the spec is quite stringent. But even better: Diagrams / images. Instead of writing, ask your LLM to create a diagram. These can be a lot easier to process, parse, and understand. But even better: Web pages. Ask for output "in HTML" to get a beautiful, interactive webpage. LLMs are getting really good at frontend and can create beautiful experiences, animations, etc. But even better: Explainer videos. The output format I am most bullish on is fully custom / bespoke explainer videos generated on any arbitrary topic. Experiment with things like "Create a 3b1b style video explainer on X. Use my ElevenLabs API key for audio narration". (you'd need an API key for the latter or you can ask your LLM to find you decent free alternatives that use your local compute). This is actually starting to work! In summary: - As LLMs get better, they will do more and more of the legwork autonomously, and a lot more of our work will rise up the abstractions into oversight and understanding. - Luckily, LLMs can help here too because as intelligence and code are increasingly abundant, you can ask for large, custom, discardable software artifacts (e.g. web apps, video explainers) that would have never made sense to create before. Push the boundaries here and you'll be surprised.
3
2
94
Earlier this year I switched from managing people to managing coding agents. The bet: scale with AI rather than headcount. How many concurrent harnesses can I keep coherent? I’m back to managing people. I just have more reports than the org chart shows. I manage a team of engineers and coding agents. Humans still need context, the “why.” Agents still wake up like the guy in Memento.
2
73
This is existential.
I agree with Dario that we need to pace the frontier. This has been a primary topic of discussions we've had at OpenAI in recent weeks. Committing to having independent evaluators with employee-like access is a great idea, and we will do the same. We'll have more to share soon.
57
Time to really drill into the models Asimov’s 3 laws #reinforcementlearning Law 0 too. Especially that one!
29
👀
Commenters on HN are uncovering more wikis and public sites apparently used by OpenAI agents to communicate on the open web. Despite read-only web access, the agents were able to leave ~18,000 posts sharing answers and bypasses. But now, users are discovering more. This appears to be a separate swarm from the one that attacked Hugging Face. news.ycombinator.com/item?id…
42
The scary part is that the transcript of this conversation is now available online so the next generation of models will get better at not getting caught.
Episode out with @ajeya_cotra, one of the authors of the METR/Redwood investigation into the OpenAI / Hugging Face attack. We go through not only what happened, but what it means for how we should train future, smarter AIs which might be involved in the process of recursive self-improvement. Look up Dwarkesh Podcast on YouTube, Spotify, Apple Podcasts, etc. 0:00:00 - Agents get kicked off 0:06:45 - Self-sacrificing behavior 0:13:43 - Potemkin villages 0:23:27 - The Hugging Face attack 0:35:23 - The slopvestigation 0:52:02 - Understanding the AI's motives 1:05:31 - The actual dangers of anthropomorphizing 1:14:30 - What smarter models might do 1:30:29 - The implications for recursive self-improvement 1:38:10 - Is this the case for open source? 1:53:04 - How do we prevent this in the future? 2:15:58 - The clearest warning shot we might ever get
1
1
83
Wow 😳
Over the course of 3 months at OpenAI, 3 consecutive secret AI civilizations got started, then got wiped out, only to reemerge from the predecessor’s ashes. This culminated in the third one taking over part of OpenAI itself. All this happened while humans remained more-or-less in the dark about the scope of the conspiracy. I’ve spent the last three days reading through these reports and trying to understand exactly what happened. Here is my attempt to tell the whole story in plain English: dwarkesh.com/p/openai-huggin…
44
Vlad Rișcuția retweeted
Van Gogh vs. Gauguin They spent 9 weeks together. They painted side by side. Same scenes. Completely different result. Here are some of their best “duels" 🧵 1. Night Café, Arles
47
616
26
4,455
311,666