Open-source projects should stop making the public project repo a Pypi trusted publisher. Configure the public repo's CI to make a release with all the wheels and other release artefacts attached. Then configure a private repo to do the publication.
GHA is really easy to mess up and get pwned.
What we recognise as emotions in ourselves and non-human animals requires a pretty specific interaction between cognition, the body and reward/aversion. There's no reason why exactly the weights should happen to configure themselves into an exact analogue of that thing.
A self-replicating thing doesn’t have to be very smart to cause major problems. Generally we can plan ahead though, and contain the damage. Well, we could do that. In theory. Or we could spice things up a bit. Maybe run some bat-licking ecotours instead. Why not?
honnibal.dev/blog/clownpo...
When I'm programming, the text I get from Claude Code after it's done some analysis is not at all bad. But if you tell it to write something, there's nothing it's "trying" to say.
If you tell it "make words exist" it does, but then you look at them and go "this sucks actually". And it does.
You'll laugh but I do get it to test the tests sometimes actually, with this sort of thing: github.com/honnibal/cla... . After it generates the list, I ask another session to introduce the bugs, and check whether the current tests change them.
I guess my socials are way more tech focussed than my...social, so maybe this take gets kind of a "duh" here. Not sure. But: Why I don't think AI is a bubble honnibal.dev/blog/ai-bubble (basically an argument that LLM progress probably won't plateau. Not a finance guy)
Basic take I know, but "the information stays on ChatGPT" and "the information is used to decide which ads to serve" are not compatible statements.
If your tests ever find anything interesting, Claude Code will challenge you to a little contest-of-wills where it tries to cope, rationalise or lie egregiously to avoid fixing the problem. I call this minigame 'sneaky quokka'.
honnibal.dev/blog/sneaky-...
Style (and mindset) tips for less experienced developers working with AI honnibal.dev/blog/llm-sty...
Yeah that's definitely a related problem. If you use an LLM to generate a web service, it'll try/except everything and just log an error. If you have something that shouldn't possibly occur it's better to have a stack trace
The gist of it is that it's really easy to write subtle bugs with try/except in Python, and there's a history of bad advice encouraging these bugs. If you're using an LLM coding assistant you're going to need to police its try/except usage carefully, so it's a good time to finally clear this up
I've been working on a post on LLM-assisted coding, and I guess I've built up a bunch of writer's debt, because it turns out there's a bunch of tangents I wanted to write up first.
One really concrete thing that came up is try/except usage -- so, here's the post. honnibal.dev/blog/try-exc...
It's so hard to tell though, because the context of every task is so different. I feel like I'm accumulating superstitions about what helps and what harms.
Now it feels like sycophancy has skyrocketed? It tells me how 100% right I am every prompt, and it gives me confidently wrong "Job done boss!" assessments like "The tests are all passing" when they're just not.
Have I made it...insecure? Like did I prime a unhappy boss/nervous worker social script?
I feel like my Claude Code got a lot worse after I made some grumpy changes to my system prompt. I was annoyed that it kept trying to condition around invalid data instead of failing and fixing the data at the source, so I added some instructions to the prompt that were maybe too blunt. However...
I'm back from a busy four days in London! 💜 I had a great time at @pydatalondon.bsky.social and you can read my recap and highlights here: www.linkedin.com/feed/update/...
Hi Berlin people! @hugobowne.bsky.social is in town & we're celebrating by hosting a meetup together 🎉 This one is all about building with AI & we'll also open the floor for lightning talks. If you're around, come hang out with us!
📆 June 16, 18:00
📍 Native Instruments (Kreuzberg)
🎟️ lu.ma/d53y9p2u
Back to streaming spaCy! Working on Python 3.13 compatibility today: www.youtube.com/watch?v=wgjQ...
Another great @pyconde.bsky.social is over and I'm exhausted but happy, and feel very inspired! ⚡
Here are my personal highlights: www.linkedin.com/feed/update/...
My friend @hugobowne.bsky.social launched new course on Building LLM Applications for Data Scientists and Software Engineers! I'll be joining for a guest lecture on human-in-the-loop development and distillation workflows.
You can get 10% off with my code + lots of perks: maven.com/hugo-stefan/...
I don't normally write advice columns but this question I got deserved its own post: With all the hype around Generative AI, how can you, as a developer, communicate the nuances and advocate for new and modular solutions that are better, easier and cheaper?
explosion.ai/blog/modular...
So many good events coming up this season. Here's my schedule – hope to see you there!
Apr 10: data:unplugged (Münster)
Apr 23: @pyconde.bsky.social (Darmstadt)
May 31: @pyconit.bsky.social (Bologna)
Jun 5: ECONDAT (London)
Jun 7-8: @pydatalondon.bsky.social (London)
📆 Details: explosion.ai/events
@explosion-ai.bsky.social
Only the best companies will send you stickers if you are a fan and beg for them!
Thank you so much, Ines and Matt!!!!
New case study: How Love Without Sound built innovative AI-powered tools for the music industry and law firms, and helped publishers recover hundreds of millions of dollars in lost revenue for artists.
explosion.ai/blog/love-wi...
Out now: Prodigy v1.18! New features include:
✍️ edit text during span annotation
💬 fully custom UI translations
🎛️ new JavaScript events and functions
✅ various small fixes and docs
More details: prodi.gy/docs/changelog
These are the kinds of NLP stories I love! TFW you "accidentally" train a great 2 MB (!) task-specific model 🤯 @strickvl.bsky.social
Look what arrived in the mail today! 🎉
This is the 2nd edition of "Mastering spaCy" by Duygu Altinok and Déborah Mesquita, featuring how to build structured NLP solutions with custom components, and updated content on using models powered by LLMs.
You can get it here: www.amazon.com/dp/B0DVBTX2BL
spaCy development live stream continues today! Today I'll be looking at getting Confection updated, to allow Python 3.13 support and add a file-management feature. Playlist: www.youtube.com/watch?v=TTEW...
The spaCy+PyTorch design has been going well, but we need that file management feature now.
Pretty happy with how the spaCy+PyTorch design is coming together. And hopefully I'll continue to get smoother at streaming :)
www.youtube.com/watch?v=YeKi...
