willj
willj on Hacker News
How can you tell if someone is a polars fan? Don’t worry, they’ll tell you. :)
Fair point. Though I don’t think time without money is really leisure time :)
It’s too bad that, yet again, instead of the productivity gains leading to shorter work weeks, the benefits accrue to the companies. Just once I’d like to see productivity gains lead to more leisure time, not higher expectation.
The temperature parameters largely went away when we moved towards reasoning models, which output lots of reasoning tokens before you get to the actual output tokens. I don’t know if it was found that reasoning works better with a higher temperature, or that having separate temperatures for reasoning vs. output wasn’t practical, but that’s my observation of the timing, anyway. And to the other commenter’s point, even a temperature of 0 is not deterministic if the batches are not invariant, which they’re not in production workloads.
If you’re using a model from a provider (not one that you’re hosting locally), greedy decoding via temperature = 0 does
not guarantee determinism. A temperature of 0 doesn’t result in the same responses every time, in part due to floating-point precision and in part to to lack of batch invariance [1]
[1] https://thinkingmachines.ai/blog/defeating-nondeterminism-in...
Thanks! That makes sense. I suppose this requires commit messages or PRs to indicate code was AI-generated vs. not, or to assume that commits after a certain time period were all from AI coding. It’d be an interesting analysis. Maybe there’s already a study out there.
In any case, thank you again!
Posted by
willj on Wednesday January 21, 2026 @08:41AM
from the ask-hn dept.
Here’s one place where I think proponents and skeptics of agentic coding tools (Claude Code, Codex, etc.) tend to talk past each other:Proponents say things like:- “I shipped feature X in days instead of weeks.”- “I could build this despite not knowing Rust / the framework / the codebase.”- “This unblocked work that would never have been prioritized.”Skeptics say things like:- “This might work...
Can you say more about the approach you take for summarization? Are the papers short enough that you just put the whole thing in the context window of the model you’re using, or do you do anything fancy? I’ve tried out various summarization approaches (hierarchical, aspect-based, incremental refinement), and am curious what you found works best for your use case.
Thanks! I got some initial ideas from Nano Banana, actually, but then spent a while iterating on different layouts myself.
This is something I built over the holidays to support people having a hard time with the short days and early sunsets:
https://sunshineoptimist.com.
For the past several years I would look up the day lengths and sunset times for my location and identify milestones like “first 5pm sunset”, “1 hour of daylight gained since the winter solstice”, etc. But that manual process also meant I was limited to sharing updates on just my location, and my friends only benefitted when I made a post. I wanted to make a site anyone could come to at any time to get an optimistic message and a milestone to look forward to.
Some features this has:
- Calculation of several possible optimistic headlines. No LLMs used here.
- Offers comparisons to the earliest sunset of the year and shortest day
- Careful consideration of optimistic messaging at all times of year, including after the summer solstice when daylight is being lost
- Static-only site, no ads or tracking. All calculations happen in the browser.
Posted by
willj on Sunday January 11, 2026 @03:00PM
from the show-hn dept.
https://sunshineoptimist.com
This is something I built over the holidays to support people having a hard time with the short days and early sunsets. For the past several years I would look up the day lengths and sunset times for my location and identify milestones like “first 5pm sunset”, “1 hour of daylight gained since the winter solstice”, etc. But that manual process also meant I was limited to sharing updates on just...
I think the models are so big that they can’t keep many old versions around because they would take away from the available GPUs they use to serve the latest models, and thereby reduce overall throughput. So they phase out older models over time. However, the major providers usually provide a time snapshot for each model, and keep the latest 2-3 available.
This reminds me a bit of using LLM frameworks like langchain, Haystack, etc., especially if you’re only using them for the chat completions or responses APIs and not doing anything fancy.
I think this ignores that the monopolies have the power to buy up any new competitors, or to drive them out of business using monopoly power. Regulatory hurdles are only one tool that (can) benefit monopolies.
I think that’s different. AlphaGo is using reinforcement learning in a context in which there is a clear evaluation function— did a strategy lead to a win or loss.
Relatedly, the OCR component relies on PyMuPDF, which has a license that requires releasing source code, which isn’t possible for most commercial applications. Is there any plan to move away from PyMuPDF, or is there a way to use an alternative?
I’d argue Bitcoin is Obscene Energy Demand.
Where are the happy offices these days? Which companies are the new “Google” who people are very eager to work for?
I like this idea, but isn’t this a recipe for only ever doing the urgent stuff, not the not-urgent-but-important stuff? For example, if your list had “read 1 chapter of SICP” on it, you might never get to it.
One thing with fastai that annoyed me when I built a project with it was that the v1 and v2 APIs are totally different, and if you google or search stackoverflow for help with something, I found it more likely to stumble on answers for the v1 API than the v2 API. I also didn’t find their documentation super helpful for more than the most basic things (though not all documentation can be as amazing as scikit-learn).
I’m not sure if you’re aware of this option already, but Google Colab sounds ideal for this situation. People get a copy of your notebook, which they can run, edit, and try new things, and all the computational resources are free. It even offers access to GPU. It’s a surprisingly robust resource. There are runtime limits, but they’re something like 12 hours, so your use case fits in there easily. Hope this helps!
I do the same, but I think there are an increasing number of astroturf campaigns taking advantage of that, too.
Can you imagine if someone had kept their crypto wallet in this thing?
Makes me wonder how the engineers will fix this if they can’t visit Stack Overflow :)
Taxing a quantity is way simpler to implement, and maybe more effective, than making a huge list of items to tax, thereby creating loopholes.
Isn’t this the same as “pay with PayPal/Amazon pay/etc”?