Ivan Milev

@yacineMTB I think understanding shouldn't be ever outsourced, however I believe there can't be an argument about thinking if you have no understanding. Is it even your work if you have no understanding of something?

you can outsource your thinking but you cannot outsource your understanding

@realsimon DMed

@ivan_milev21 Hey, let's chat if you're raising funds. Send me a dm.

Showing off my unconference merge while explaining why we should not turn our codebases into blackboxes. Thanks for the free shirt 😎😎 @dexhorthy (@humanlayer_dev) and @vaibcode (@boundaryML) New release is dropping this week (@CodeBoarding)

The @aiDotEngineer has been dropping pure gold on YouTube. Been glazing hard on these: youtu.be/v4F1gFy-hqg and youtu.be/RjfbvDXpFls. Love seeing more visibility for the fact that good engineering matters more than ever, LLMs are pure GIGO. Thanks @mattpocockuk @badlogicgames

@art_zucker @IlysMoutawwakil W/o MoE local LLMs are dead imo :D (I want to code on airplane mode maaan)

Today I re-iterate: I hate MoEs and we are wasting time on them.... Let's unite and call a global ban on MoEs please. Please 1M+ salary researchers: do better... credits to @IlysMoutawwakil for the graph:

@GithubProjects github.com/ivanmilevtues, followers on GitHub are 100x more impressive than those on X, change my mind.

|￣￣￣￣￣￣￣￣￣￣￣￣￣| | Share your GitHub profile. | |＿＿＿＿＿＿＿＿＿＿＿＿＿| \ (•◡•) / \ / —— | | |_ |_

@paulg It feels like being non-AI is now a differentiator and somehow a defensibility angle, as OpenAI and Anthropic are less likely to kill you on a tuesday morning. Tbh most ppl have forgotten that one doesn't need an LLM to rename a variable or to classify a spammy email...

The biggest opportunity for would-be startup founders is AI. But the most underpriced opportunity is probably non-AI ideas. So if you have a good non-AI idea, go for it, because everyone else is going to overlook it.

@dexhorthy Super well said, SWE is much more than coding and momentarily output. On mentorship, I feel the same about code reviews. As a junior, I learned the system by reviewing others’ work, understanding their thinking, and discussing tradeoffs (on the PR). Man, I had a great team.

you don’t hire interns because they’re gonna ship mad features Every intern is an investment in a future hire. Perhaps A modern day apprenticeship. You don’t hire juniors cause you need that extra 2-3 bug fixes per day. You hire juniors so you can turn them into seniors When you say “ai will replace juniors” all you’re telling me is you don’t understand mentorship in software careers, or you think of every engineer as a ticket factory. We do need to rethink how we upskill this new generation of SWEs though. The best engineers are great at AI (coding, context eng, etc) and great at software engineering - (systems, algorithms, debugging, architecture) More seniors are good at the latter, more juniors are good at the former, but a lot of ai coding takes away the “friction” which, as @badlogicgames so helpfully pointed out, is where you learn

@dexhorthy I still have a strong believe that such projects dont really work at high scale, outside of green field projects. But they will open new job positions like vibe code cleaner, vibe code explainer, vibe code consultant

in general i'm not opposed to the goals of gstack i'm not opposed to yc promoting gstack as a general lesson in "use skills and import other people's knowledge" if I have any feedback on gstack it's at the technical level and optimization opptys on how transformer attention works (and I tried to ping garry to share some of these thoughts but he's a busy man) - instruction budget - many of the gstack commands have well over 200 instructions. That means limited adherence to all of them, and would be better off routing to smaller more focused skills (see link in comments for Monday's @aiDotEngineer talk on the subject) - instruction budget - most skills have long preambles about setup / init / upgrade steps, long inline bash scripts that get read every time, do not follow progressive disclosure at all - inline bash and prompting to do state management and version updates etc - clever, almost novel, as @davis7 pointed out this is building large complex programs in markdown, fantastically futuristic. But again, including it in every skill invocation means you're wasting a ton of tokens and attention on noise instead of putting attention into solving the user's problem with *their* context all that means that as well-meaning as the system is, and as bullish as I am on "the YC way" and packaging that knowledge and giving it to everyone, it could be SO SO SOOOO much better on the execution side

@Tocelot @speedrun You should take a look at @CodeBoarding we've been on this for few months now. Coding agents in CLI are just a blackbox that wrecks your codebase, we generate a higher-level representation so one can see what the heck is going on

a16z @speedrun request for startups: GUIs for Agents we’re still in the MS-DOS era of agents today - CLI, terminal sessions, file directories deleted by openclaw etc. while a small slice of silicon valley are power users, we're SO early for the rest of the world at Speedrun, we’re looking for bold founders excited to bring the power of agents to normies everywhere. there's a whole slew of products to be built here - from agent builders to marketplaces to managed infrastructure one broad idea we’re excited about are visual abstraction layers for agents. if you don't know exactly what you want, a command line / chat interface is paralyzing - you need to see options 1 example - think of a GUI or visual command center inspired by strategy games (ex. Factorio) where agents and workflows are represented graphically. skills, tools, MCP connections, background processes, etc could all be configured and shown visually in a workspace on UX, strategy games have long perfected agent management. zoom to get a birds-eye view of your agents, batch and queue orders via shortcuts, assign agents in multiplayer etc. a well-designed agent command center would make multi-agent orchestration for normies feel easy & intuitive most folks today still haven't moved beyond ChatGPT. the potential is enormous - just as Windows unlocked mass-market use of personal computers, the right visual abstraction layer could unlock agentic work for everyone - from individuals to enterprise teams if you share our vision, we'd love to chat!

Everyone talks about how coding agents brought excitement to coding as you can do everything in a day. No one talks about how they took satisfaction away, there is something super satisfying in writting up a tight implementation, somehow a tight plan doesn't hit the same thou...

@lolesports @Thorin Bro is locked in, doesn't outsource anything to AI, fully present

OUT OF THIS WORLD.

@thdxr Read your tweets...

so what do you do when your agent is working and don't say start another agent because i know you're lying

@LuthraAbhyuday Cars is absolutely my favorite cartoon of all time (a close second is Kung Fu Panda). “Focus. Speed. I am speed.”

“i just never thought i couldn’t” - lightning mcqueen

I think that at 90+ percent of the events I've been to in SF people are mentioning PR and Lines of Code ae measurement for SWE productivity. Cant wait for people to find out about ast fuzzers, productivity will go through the roof...

@dexhorthy @barry_zyj @MaheshMurag Curious to see if we end up doing spec driven development. Then we'd build up specs on different levels of detail instead of skills (I would imagine)

"thick skills thin harness" is basically getting at @barry_zyj and @MaheshMurag said back in november "don't build agents, build skills instead" - youtube.com/watch?v=CEvIs9… and while "put the biz logic in the skills + the deterministic code they reference (CLIs)" is probably +EV my concern is that the particular phrase "fat skills" will encourage our friends to build long skill files with too many instructions and you have an instruction budget. youtube.com/watch?v=YwZR6t… even the best frontier models have degrading performance the more instructions you try to stuff in. you always get more intelligence and better results if you limit the number of instructions. Thin harness, thin skills, skilled operator. x.com/garrytan/statu…

@thdxr I feel like email communication is slowly shutting down because of LLM automated campaigns and spam...

AI safety focuses on dramatic things like bioweapons and nukes but LLMs are breaking so many systems we rely on in more boring ways via spam, noise and misinformation wish this was more of a focus

@corbin_braun Im around town, lmk about details

who in SF is building something cool and wants to hop on a podcast today in person?

Just noticed that it is free advertisement for @opencode (love it), we should put something similar for @CodeBoarding as well hahah

POV: Showing what agent coding should look like to @brycent . Great meeting you at the @UseCorgi cafe man. I feel like I might start streaming some coding sessions on tw after chatting with the guy in front of the camera...