Alexander Kitov
1.9K posts
Alexander Kitov
@_akitov
VP Product @TelerikAcademy. 11+ years as Product Manager and Entrepreneur. Father, continuous learner, AI-optimist, and tennis enthusiast.
Bulgaria Katılım Nisan 2010
719 Takip Edilen622 Takipçiler
@_akitov @danshipper @cursor_ai no alt tab is better, less friction is good, and lookign or interacting with the app can mean you queu up next steps
English
I am more and more convinced that this is the future of software development UI. @cursor_ai is the closest in my opinion
A list of work you're working on parallel, the agent in the middle, and most importantly, the thing you're building on the right. Because you want to see what you're building, you want to give feedback, you want to work on it. As a builder, the code is not important anymore. As a builder, it is all about what you're building. So seeing the app next to the agent is very important.
English
@kieranklaassen @danshipper @cursor_ai if interacting magically turns into queued prompts or tasks, that would be a good reason to force the app into the env
having used replit for a few months, i never used the preview, but corsor might have done it diff
will give it a try with a responsive web app and report back
English
I jus took the first PNL screenshot since the uptrend started
English
Alexander Kitov retweetledi
AMD Senior AI Director confirms Claude has been nerfed. She analyzed Claude's session logs from Janurary to March:
> median thinking dropped from ~2,200 to ~600 chars
> API requests went up 80x from Feb to Mar. less thinking and failed attempts meaning more retries, burning more tokens, and spending more on tokens
> reads-per-edit dropped from 6.6x → 2.0x. model stops researching code before touching it.
> model tried to bail out or ask "should i continue" 173 times in 17 days (0 times before March 8).
> self-contradiction in reasoning ("oh wait, actually...") tripled.
> conventions like CLAUDE.md get ignored because there's less thinking budget to cross-check edits
> 5pm and 7pm PST are the worst hours, late night is significantly better. this means the thinking allocation is most likely GPU-load-sensitive.
English
bruh u guys remember these
I think they had like bootleg tetris, bootleg racing, and a bunch of subpar games you’d only play if you were beyond bored
Kalbur Saman@kalbursamantr
Gerçek araba yarışı buydu.
English
To simplify our system, I asked Opus to port the file indexer to native TypeScript. Here's the prompt I used:
> compile our rust+napi file index to native ts. compile the code over, and make sure the code perfectly passes the ORIGINAL test suite from the package. then profile compared to napi/wasm, and put up a pr with the result. dont stop until you have conclusively proven that the new impl is faster.
I then followed up with:
> anything else you can do to improve performance? profile it in . hoping for p99 < 10ms
And:
> im still feeling a little stutter when i first start typing "@ cc"
Finally:
> nice that feels really good. /go
English
Just got a nice DM from a big enterprise customer using Claude Code in one of the world's biggest codebases
Here's how we made @-mentions 3x faster in large enterprise codebases 🧵
English
@tobi so qmd is already operational to run in a cloud environment?
English
@tobi @thorstenball it’s the inner fight with oneself, given the quickly changing circumstances of the gameplay, all happening one point at a time
clearheadedness, introspection and mental reset before and after each point being mandatory
it’s a fascinating game
English
@thorstenball Insanely good book. And I feel like he totally didn’t realize how good this book was while writing it either.
English
This was amazing.
I don't play tennis, have no real interest in it, but this was one of the best things I've ever read on learning, improving one's performance (whatever you want that to mean), focus & attention.
English
Every fortune 500 company right now is spending insane amounts of money to figure out how they can use AI with a harnass that guarantees their security, alignment, access levels, protection of IP, corporate strategy and data etc. Governance, done right, is the key to those guarantees. I imagine that AI will become one of the largest expenditures on anyone’s balance sheet and the harnass to execute this will be leading the rate of adoption, because trust is the key.
@thestorecloud is solving for this.
English
@Vtrivedy10 don’t get why always on heuristics and task specific shouldn’t exist in parralel
but i get the task specific harness being a step change for the quality and success ratio of the task itself
English
the future is probably more just-in-time agent+harness assembly based on task context
we’re just not perfect at that yet so we use a lot of coarse heuristics like “always on”
James Long@jlongster
regarding agent memory, I'm realizing: I never want anything loaded automatically. no loading yesterday's memories etc (AGENTS.md is different, not memory) I don't want chronological memories. I want topical memories grouped based on what I'm doing I want to be explicit about it generally, especially saving but also reading. some automatic reading based on me switching "into" a topic might make sense most thing I see don't really fit what I want
English
Boris, Anthony, you guys can't keep up with everything I'm sure, but I'm planting a seed here.
Loading up large skill files (or other resources) should be rule-based, but that's possible only if the est. token size is known upfront, ie. frontmatter'ed
CC @bcherny @amorriscode
English
@mitsuhiko Armin, what’s an example for a practical use case for durable execution, there are probably many, but what’s one (in non-technical lingo) you think of immediately?
English
Since some asked: five months of Absurd in production. Durable execution on plain Postgres with tiny SDKs. The design held up fine, partitioning is still hard and some closing thoughts on the future of Open Source. lucumr.pocoo.org/2026/4/4/absur…
English
gpt is substantially dumber than claude in a 🦞 ai assistant context. it's absolutely maddening talking to this thing now.
English
@Vtrivedy10 @sydneyrunkle @sydneyrunkle Sidney, hi, are there “non coding” OS harnesses that you maybe know of? it seems a harness is currently used only in conjunction with coding agents
English
Sydney continues to drop actual harness eng bangers with code snippets every day
it's honestly the little things that compound in tasteful harness design
Sydney Runkle@sydneyrunkle
harness eng day 5: toolsets some tools need setup and teardown around the agent loop, like connecting to a tool server or spinning up a sandbox for example, @langchain's ShellToolMiddleware handles init and cleanup, and injects the shell tool into your agent's tool registry!
English
