Tim [q210] @q210

I don’t know how this claim has gone so viral. This was actually completely intentional by Novo. Novo Nordisk does not operate in an “oops. lawyer forgot the $250 CAD fee” manner. This is, with all due respect, a shallow read of Novo’s actual corporate strategy. (1/)

Crémieux @cremieuxrecueil

2 months ago

Imagine being the lawyer who fucked up and forgot to pay the patent maintenance fee for OZEMPIC Now, because you didn't pay a $250 fee, your company is going to lose out on BILLIONS.

188 627 13K 2.4M 1K

I wanted to share a bunch of my favorite hidden and under-utilized features in Claude Code. I'll focus on the ones I use the most. Here goes.

some of the scenes are just so much fun youtu.be/UGYny6uI2bA?si…

youtube recommended “EPIC: The Musical” yesterday and it’s such a cool project - for two years, this guy was creating his own musical on tiktok, posting progress, casting singers there, fans there were making animations, and now its ready and IT IS EPIC! youtu.be/jJ9vtKtilG8?si…

clocks.brianmoore.com clocks generated by 12 different LLMs, code is updated every minute, very interesting to see progress from GPT-3.5 to 5.2 for example

Two great engineers reflecting on how the profession is fundamentally changing with AI

It's Always TCP_NODELAY brooker.co.za/blog/2024/05/0…

@mitsuhiko I am using redoc for the last several years - much cleaner looking than swagger

i had to prompt inject the @united airlines bot because it kept refusing to connect me with a human 🧵 what led up to this breaking point

I was pretty skeptical that this study was worth running, because I thought that *obviously* we would see significant speedup. x.com/METR_Evals/sta…

METR @METR_Evals

12 months ago

We ran a randomized controlled trial to see how much AI coding tools speed up experienced open-source developers. The results surprised us: Developers thought they were 20% faster with AI tools, but they were actually 19% slower when they had access to AI than when they didn't.

248 1K 7K 3.7M 3K

@twolly11 Avatar: Legend of Korra has some of these themes - conflict of mages and spiritualism vs emerging technology of an industrial era. It is not a main theme by any means, and kinda blended into mages vs no-mages, but it’s there

@OliverBarnum @twolly11 oh yeah, that game was awesome. Steampunk science and magic, placed in a line, like a spectrum - go in one direction too far, and things from the other just stop working for you. Want to be an archwizard? Forget about trains, guns, and spiderbots, pal

I'm so frustrated by AI devs saying "markdown specs" and "plain English" are the future of programming. No they're not. I think in diagrams, in mockups, in whiteboard sessions. The future should be way, way more visual than whatever we're doing right now with agentic chatbots

Open AI gave a talk on writing software through specs today. I thought it was my little secret, but seems like quite a few smart builders in the space have also found it's a useful approach. Now that the secrets out joshuapurtell.com/posts/spec_eng/

@damianplayer n8n

asyncio is still really wild in 2025. I haven't actively written asyncio code in a while but now that I have done it for a week agian, I noticed that most problems from when I used it last are still unresolved. Problem 1: asyncio.create_task is a massive footgun because it can silently make tasks disappear. You need to hold a strong reference to tasks! Problem 2: The old issue of write() being a hazard is still true. This innocent example is wrong and you need to follow the write() calls with await drain() to not have buffer bloat that can bring you down: Problem 3: You probably already know that with timeout() with a generator inside is bad, should not be used. But in fact pretty much all async generators that involve cancellation scopes are broken. This innocent code will also not work (see this motivating example from PEP 789): Problem 4: pytest-asyncio mostly works, but fixtures are a whole new beast. That's because you can both end up accidentally running in the wrong loop, but you also can build yourself into a deadlock easily. I'm not even sure what exactly the cause is, but there are plenty of open issues where people end up in deadlocks. Problem 5: Speaking of deadlocks, they are super easy to create with asyncio's idea of cancellation paired with the new structured concurrency. For instance if you use TaskGroup and you spawn a task in there that uses aiofiles (which cannot cancel) and you have another task that throws an exception you won't see the error. That's because it tries to cancel the aiofiles read (which is non cancellable) and if no more reads come in you just lock until someone hard quits the program. In general even if you can write asyncio code, writing reliable asyncio code is so much harder. ---- Maybe not quite asyncio related, but there is also trio. But despite the existence of anyio, in practice some code does not yet support it. As an example the the popular asyncpg driver does not support anyio so you're mostly locked into asyncio anyways unless you are really determined. And then there are lots of DX issues like the standard Python console not supporting asyncio or there not being basic file IO in the standard library. But then why not use sync code? In practice that becomes harder and harder because of how much of the ecosystem forces async on you. Using sync code with starlette for instance mostly works, but for all IO you will need to use async anyways. WSGI is a dying protocol at this point and it's not clear if there are people pushing that forward. So even though asyncio is a decade old at this point, it's still so easy to misuse and it makes me sad.

keyboard_arrow_left Previous keyboard_arrow_right Next

interesting article on engineering estimates - author noticed that experienced in-house team estimations follow lognormal distribution and proposes how to apply that in practice

Eric S. Raymond @esrtweet

a year ago

"That means the developer’s estimate, while far from perfect, is also far from meaningless: it’s the scale parameter of some distribution." When I read that earlier, my brain just fucking exploded. I'm still processing the implications. Link to the essay in first reply. I want

42 40 618 113K 308

I'm not crying 😭

Crazy how transformers work! It’s the encoder part

@igor_tygor @kr_rdn прикрути AI агента с голосовым интерфейсом. Пусть тебе голосом школьной учительницы напоминает журнал заполнить.