Claude Sonnet 4.6 telling me that Claude Sonnet 4.6 doesn't exist:
"At the time of writing, Anthropic's naming convention uses names like claude-sonnet-4-5 or claude-opus-4. The 4.6 version doesn't appear to exist."
"Let me try a couple othersources"
GLM-5
I love seeing LLM slip-ups like this, today from Gemini:
"This gives a scent of what else is there..."
Imagine how grammatical, typographic, and other language errors are happily baked into the weights.
It’s very cool. If you haven’t read mariozechner.at/posts/2025-1... yet it’s a great window into how they built and think about it
The magnitude of the FAFO moment coming for the "we don't read any of the AI code we ship to prod!" crowd is going to be a sight to behold. Unfortunately, we are all going to be watching it as affected users of these systems
Why AI Swarms Cannot Build Architecture https://lobste.rs/s/uhtsz9 #scaling #practices #ai
Documentary about Open Source in Ukraine and around the world https://lobste.rs/s/unmf9w #video #culture
Really enjoyed this take on Carney’s speech. ⬇️
This article has been bouncing around in my head for the past couple of months, but I finally sat down and wrote it.
I can't believe how often I hit this ridiculous rate limit just by browsing to things on GitHub's web site--many times this week.
The downfall and neglect of GitHub is hard to watch and brings me no joy.
I missed this piece by @dkthomp.bsky.social which makes a much deeper case for a thing I said to @blaine.bsky.social over the summer.
Instagram shouldn't be understood as falling to be a social sharing app. It should be seen as a very successful Pocket TV.
www.derekthompson.org/p/why-everyt...
I'm grading my open source students' work, and I loved this description of what it's like from one of the students:
dev.to/jongwan93/es...
I listened to a podcast with the guy who wrote the extension and it made me think that so much more is possible than I was thinking before
I was reading the source for this last week, it’s really well done
Seeing a lot of what I'm calling "AI Ennui" with my students. Different from the usual end-of-term burnout and apathy, it's become increasingly hard to get students to be students (read, write, engage with learning). Blaming AI is too easy, but it's fundamentally different from anything I've seen.
Wild to see upgrade docs being given as a prompt for your LLM agent to do...
www.prisma.io/docs/ai/prom...
Another day, another reason to love OpenRouter.ai. Student API key leaked into a public GitHub repo and we get a security email from their support team, which have already found it and disabled the key.
Fantastic
I never cease to be underwhelmed with the average case for AI image generation. Everyone showing images of Nano Banana generating whiteboard images of papers.
I try and it fails to do anything 3 times, then creates one about random AI stuff vs. the biology in the paper itself. Cool but useless?
What is this nonsense in Chrome?
I read about this a while back too and was blown away that the briefcase icon I’d ignored for years was able to sync!? Amazing how not well understood it was.
I continue to bet on OpenRouter.ai
Today I'm writing my LLM programming course notes on using embeddings and they add embedding models: openrouter.ai/models?fmt=c...
Then I get an email that they're doing their own TS/Python SDK: github.com/OpenRouterTe...
Every week they ship great stuff
Remember IBM’s Watson?
Using them for a PWA to replace SMS notifications, really loving so far.
Things that surprised me with Apple’s impl:
- no image support (my users expect this and MMS did it)
- unable to send multiple in quick succession without some being dropped. Not clear what that threshold is
Interesting Claude Sonnet 4.5 hiccup, where it complains about bits of its own system prompt when reviewing a document I share
"Token budget directive at top: The <budget:token_budget>1000000</budget:token_budget> appears to be a system prompt artifact that shouldn't be in student-facing materials"
Great post. I also find it interesting to watch how newer React also can’t overtake older React
Upload your scanned ID they said. It will be fine they said.
