nate
Engineer at TLDR
- Joined
- Nov 17, 2026
Comments
Why are all the indie-friendly products dying at the same time?
One thing I remind myself of a lot is that so many people don't make spreadsheets even when a spreadsheet would be really useful. It's getting a lot easier to make software, but there's some sort of upper bound (maybe not a hard upper bound, but you know) of how many people are actually going to make themselves software that's suggested by the proportion of people who will actually make themselves a custom spreadsheet.
Why are all the indie-friendly products dying at the same time?
I loved this write-up! I have such warm memories of CSS Tricks, because Chris Coyier's podcast was one of the first ones I listened to every week when I was first getting into software engineering more seriously. It's so great to see these institutions like CSS-Tricks survive.
Switching LLMs mid-conversation is a ski-rental problem - Edgee Blog
Totally! Thanks! FWIW Pangram flags this as majority AI, and it reads as AI to me. I don't at all doubt the care that went into the piece, but when it comes to things like governing metaphors and high-judgment discussion of where the tricky parts of an implementation are, I'm eager to get it more directly from the human.
Again, this is very cool and I'm glad to have come across it! I'll write to him on LinkedIn now.
Maintainers don’t need more AI bug reports. They need fixes.
Yeah, this makes me happy to read. It's going to be pretty tough to shift our collective habits, but there's just so much less reason than there used to be to submit an unverified bug report (or, in many cases, a bug report without a suggested fix). This is both because (i) verification and remediation is so much cheaper (on average) than it used to be and (ii) the value of saying "hey, problem X exists" is so much lower than it used to be (because re-discovering it is so much easier).
Switching LLMs mid-conversation is a ski-rental problem - Edgee Blog
I really had not appreciated how subtle this problem is; I like how the introduction exposes my ignorance to me so efficiently! I'd love to read a human-written follow-up that goes through some examples and discusses some of the implementation details from a different perspective.
[Weekly] What are you building this week?
I was just doing this yesterday! Dictating notes is a good way to start blog posts that I don't quite know how to frame. And playing with d1.2 has made me realize just how valuable the last X% improvement of audio-adjacent AI has been and how valuable the next X% will be.
Introducing GPT-6.2 Astra and Sol
Not that exact effect, but spiritually analogous ones, definitely. So, for example, slightly improving the information we send to decision models tends to make them vastly outperform slow LLMs that don't have the better information.
This is slightly speculative / impressionistic, but I think this effect is even bigger in the real world than the numbers would lead one to believe, just because benchmarks are less and less useful (natemeyvis.com).
Introducing GPT-6.2 Astra and Sol
Agree! FWIW I didn't explain myself clearly enough to show why this wouldn't be quite so cheap: these are eval harnesses, where Opus was involved in finding and drafting the evals, but humans have reviewed, edited, and modified a lot of them. The "Opus-free" version of this is pretty labor-intensive because of the searching and human effort required (and the harness is already fractionally non-Opus, because I use a bunch of models).
This also means that it's relatively unlikely (IMO!) that much of the difference here is Opus bias, but "relatively" is doing a lot of work there.
Also FWIW, Opus absolutely smoked Astra this morning at another qualitative task ("what's the best way for this system to handle X?").
[Weekly] What are you building this week?
This is a fun combination of "yeah, makes sense this would exist now" and "holy cow it's incredible that this exists."
[Weekly] What are you building this week?
I'm curious about the activity logger! It reminds me a bit of a project of mine that (i) started pre-AI, (ii) is a sort of practical everyday use kind of application, (iii) came to be because I wanted one of its type with a certain big custom feature set, and (iv) is very amenable to piecewise enhancement with / replacement by AI, though it's not at all trivial to figure out which pieces, and how. It's a very interesting category of application to me, largely for technical reasons but also for emotional ones (they're often passion projects, and grafting AI onto a passion project one started without any idea that AI would be what it is can feel strange).
I'm looking forward to reading more about this project as you choose to write about it.
Introducing GPT-6.2 Astra and Sol
Just following up: I've run a whole bunch of more quantitative assessments of Astra 6.2, and it's way behind Opus 5.5 on those. Now, some of those assessments were constructed in part by Opus 5.5, but I keep a careful eye on them, and it's really not close right now. (I've tried a couple different thinking levels for Astra 6.2, FWIW.)
[Weekly] What are you building this week?
I'm working on TLDR Reader (kingfisher.tldr.fm), a place for fresh tech news.
Feel free to browse or search (you can use queries of arbitrary length). You can make one-off searches or use them to tune your feed. The stories are expandable, so that you can get the most important information about a story without clicking through. We want it to feel like a TLDR newsletter, but more customizable and even fresher.
All comments welcome! As with this forum, everything from bug reports to interface details would be great to hear about. I'll be watching this thread closely.
It's not surprising but is very cool that there's a new decision model that supports audio input. The five new question types that return more variable kinds of output (e.g., many bounding boxes and every passage of text of a type) are a bit more surprising to me, and very exciting. There seem to be no barriers to a world where there are custom questions to handle many, many kinds of use cases at decision-model-level cost and latency.
I haven't tested those new question types yet but have done some preliminary testing of audio, and in a fairly challenging domain: drilling trivia, where there's lots of "uh" and "um," niche vocabulary, proper names, and non-English words. d1.2 has been impressive so far. I plan to exercise this use case a lot more.
The Gamma Model Index: OpenAI takes the lead for the first time
Misc. reactions:
This is cool and I'd love to see more analyses like this that combine qualitative and quantitative description and describe performance in terms of roles like "editor."
I strongly agree with the note that open models' performance on benchmarks is far ahead of their actual utility. I've found this to be true just about every time I've tested one.
What I like about this piece (the mix of qualitative and quantitative description of real-world usage) means, I think, that it really ought to have been written entirely by a human. Pangram and I agree that it was not, and I think it's important to interpret the various conclusions as AI-guided.
Introducing GPT-6.2 Astra and Sol
I'm pleasantly surprised by the speed.
Opus 5.5 is so good that this is the least motivated I've been in a long time to try out a new flagship model.
On my "impressionistic benchmarks" (standard routines I have for finding high-priority problems and similar), it's comparable to Opus but not clearly better. Its choice for "most urgent problem" is not good, and is the same choice I get from Opus 5.5 and Astra 6.1.