Dagens Vibes — 29. juni 2026

Theo - t3.gg

@theo

I’d estimate we’re ~6 months from most devs moving their code agents off of their laptops

@levelsio

@levelsio

✨ I think I've been coding almost solely on my VPS with Claude Code for almost a year now All I can say it's just fantastic: - no need to keep laptop open ever - no laptop battery drain - can switch to phone or any other device you like whenever you want to continue (like when you're outside) - it just keeps going all night while you sleep (esp with /goal) - you can start hacky projects from scratch and go live in seconds because you're already on the server which is great to ship things and get it used by people fast (not stuck on your local laptop webserver) - it just feels like living in the future I used to code on my laptop, test locally, then push to GitHub, then it auto pulled and deploy to production, that'd take me ~1 minute to get a new feature out But then when I bought a new Mac Book Pro a few years ago I was too lazy to install a local Nginx environment, so I just started pushing to prod and everything went fine, and I sped up deploying to about 3 seconds from laptop to server, which people called me crazy for too But now with Claude Code on my VPS in the last year, it just live edits on my production server, which sounds like it should go wrong but it just doesn't, it's very careful and only twice in 12 months messed up which meant my site didn't load for 10 seconds which is OK If I wasn't working solo, like at a big company, I' think I'd recommend the same workflow but with a staging server, so it wouldn't touch production, for safety and regulatory reasons etc. but for me it's fine I agree with @theo completely, it's clear to me this is where it's going, also seeing @karpathy with Claude moving to the cloud (via Slack etc), I think AI "agents" and AI coding will operate on servers / from the cloud first P.S. I have 3-2-1 backups, multiple on-site and off-site backups which you should also even if you wouldn't code with AI, safety first!

Gergely Orosz

@GergelyOrosz

I talked with a few folks inside Anthropic and I am starting to understand what @karpathy is saying (and what lots of people are misunderstanding) It's not about Slack, but about a cloud AI, hooked up to ALL internal company systems, that "just works." THIS is the breakthrough

clem 🤗

@ClementDelangue

the future of AI is multi-model (including a majority of open-source ones provided by @huggingface of course!)!

Ryan Carson

@ryancarson

This post is blowing up. Many people: “how can you possibly spend this much?! Just use multiple ___ Max plans for $200/mo” We’re shipping ~30 PRs daily because @HelloUntangle is growing very quickly. We have a complete software factory on @DevinAI that allows this speed - primarily because it’s cloud-first vs local dev, has visual testing out-of-the box, automated landing with thorough code review-fix loop. This allows for much more effective parallelism. FYI, we get $5k/mo in credits but we’re still spending another $15k on top. We’ve actually increased this now and are spending $1k/day for my work. I think all CEOs will eventually accept that it’s going to cost $2-5k per month per engineer (and eventually all employees because everyone will be using agents for most of their work). So the agent labs will figure out how to deliver the same number of high quality tokens but for 1/4 the price of frontier-only. The labs will probably cave as well and realize they need to do proper model routing based on the goal of the thread.

Matt Van Horn

@mvanhorn

TL;DR of my new article: I cut my agent's memory from 218 files to 6. The loud advice is "build a bigger memory system." I went the other way. Your memory and your instruction file (CLAUDE.md, AGENTS.md, whatever your agent loads) are the same problem: standing instructions that load every session and rot the moment you stop looking. 🧠 A fat memory makes your agent dumber. It injects every session, and a full context window performs worse. Keep each file under 200 lines 🗑️ Most "memory" is misfiled: trash (git already has it) or a skill lesson sitting in the wrong place 🛠️ Anything you do twice, write a skill. A lesson belongs in the skill as a PR, where it helps everyone, not journaled in memory where it rots 💾 Back up the whole store first, then cut to the bone. Deletion is only fearless when it's reversible 🔌 Turn auto-memory off. Move your survivors into the always-on instruction file so they still load 📄 Project instructions are one line: import the shared file. Write conventions once, every agent reads them (AGENTS.md is the cross-tool standard) ✂️ Stop using your instruction file like a trash can. Delete anything the model can infer from the repo. Every line earns its slot or gets cut 🔎 Still want recall? That's pull memory, not push. A brain you query, not one you carry into every session (gbrain, supermemory) Your instruction file is the most expensive real estate your agent reads. Pay for signal, not square footage.

Bartosz Naskręcki

@nasqret

I am posting much less in the last few weeks. This is mostly because I am currently doing a lot of very fun mathematical projects, all of them enhanced with the auto-research methodology. One very happy story with @PiotrPePo is announced in the illustration, the preprint is coming soon. We worked on this project the last three years and recently disproved the slope conjecture, this is a major step in the field. The paper, new insights, computations and the AI contribution will be described soon. There are some big surprises. Stay tuned! More about auto-research and my recent methodology. By auto-research I mean a scaffolded workflow with agents, coding tools, literature search, computations, memory files, and human steering. Not autonomous discovery. More like building an interactive research landscape around a problem. The scheme is the following: 1. I focus on the general problem, existing literature, known edge cases, conjectures, and timelines. 2. I build a targeted line of auto-research to dig deep into the topic. This is usually not meant to be a discovery line. It is a focused run in which I want to see which parts of the problem are cheap to establish, which questions are naive, which are computational, which are conceptual, and which are potentially hard research quests. 3. I build a very detailed roadmap for myself to see how deep and far people have reached into the problem. The auto-research part here is absolutely brilliant. It allows you to make a high-resolution panorama of the topic. I am not focusing yet on new contributions. It is like a sketch which allows me to "see" the problem. This has never been possible before at such a scale. 4. Once I have this "topographical map", I can start the proper research part. This is where deep insights, the heaviest models, new tactics, papers, and very heavy server computations enter. 5. The control of such a project is now very interactive. I basically have a responsive scaffold. I can generate my own insights, my collaborators' insights, write drafts, accumulate knowledge in knowledge bases, and absorb any scrap of knowledge I can possibly imagine. 6. In this phase, I switch very actively between human reading, highest-tier model screening, and sheer lucky guesses. 7. Once I manage to establish something very novel or promising, I decide either to focus the project on this particular insight or to divide the work into subtasks and small goals, and make a detailed list with an evaluation of which goals are more feasible than the others. In my toolbox, I currently mostly use Codex CLI, occasionally Claude Code, all of them with top-tier reasoning models, including some experimental and future releases. I use a substantial amount of tokens, but my projects are not based on gigantic repos, so the token spending is mostly due to reasoning effort focused on the mathematical problems. The results are at the moment mixed. I do see a lot of progress in problems based on heavy computations, coding, and those which looked hard because I could not put all the puzzle pieces together. There is no AGI-like model yet which can do all of the research by itself. The models need at least a very solid scaffold and original ideation; they also feed heavily on the existing literature. But I see glimpses of self-correcting research loops. This is the most remarkable observation. Once the setup is correct and I feel that we are close to the solution, auto-research is usually able to find a complete solution to the task, and my intervention becomes very moderate. Writing down a human-understandable report is another task. Here is where the human operator is very critical, because the models in such wide explorations have the tendency to build idiosyncratic terminology, odd-looking formulations of known behaviour, and to explore frantically non-promising branches of the research. In my experience, the reporting part requires heavy trimming, multi-level summaries, JupyterBooks, memory files, journal files, and knowledge-base summarizations. At this point, my style, or the style of the domain experts, must be used to carve a digestible report. Most of the otherworldly-looking reports happen exactly at this place. Repositories contain a lot of structured data, but someone has to decide what is worth writing and what the big picture is. You can even run models with heavy steering to sieve out the weeds and leave the essence. I am not sure yet if that step can be heavily automated. Many aspects of this summarization depend on taste, judgment, narrative imperative, and genuine storytelling of the domain. What is the big picture? What were people asking about? What might be interesting in a curiosity-driven fashion? I think AI systems are still very bad at this step. So this is my current view: the frontier is not autonomous mathematical discovery yet. The frontier is a new kind of interactive research scaffold, where models can explore, compute, organize, and sometimes close the loop, but where human taste and judgment still decide what the work means.

antirez

@antirez

GLM 5.2, Q2_K routed experts (effectively ~2.6 bits) running with SSD streaming on an M5 Max 128GB computer. https://t.co/Vka96HMLnS

Mark Gurman

@markgurman

There’s no question that ChatGPT has created value. But the costs are now becoming real to many customers for the first time: pricier devices, tighter component availability and an industry now shaped around serving AI infrastructure rather than making products cheaper & better.

@levelsio

@levelsio

📻 Made a Nullsoft SHOUTcast server at https://t.co/ezfSY9zJoo SHOUTcast was Winamp's own streaming platform and would let you create your own self-hosted radio stations You'd then submit it to their directory to which would get you listeners Around 2011 it reached its peak with 900,000 concurrent listeners with about 45,000 active radio stations Winamp had been acquired by AOL in 1999 for $80M and was essentially fully neglected since so it was kind of a miracle it kept going for so long But the tech is still there and you can host a server!