HNHacker News
TopNewBestAskShowJobs

radarsat1

4,157 karma · joined May 28, 2011

Info: https://sinclairs.gitlab.io

Email: <myusername>@gmail.com

submissionscomments
radarsat1··on Qodo CLI agent scores 71.2% on SWE-bench Verified
I was thinking about this recently with respect to how many agent systems now let you specify a smaller/faster model for easier tasks and a bigger model for harder tasks.

It's interesting to think about what the trade-offs are. Assuming the system can properly classify a task as easy or hard (big "if" but I guess there are ways), there is nonetheless more to think about, depending on your pricing plan.

For subscription pricing, I guess you don't really care which model runs and in fact it's hard to find a reason to ever run the smaller model, so choosing between the models is more in the provider's interests for cost efficiency.

But for pay-per-use pricing, But if you have a bigger model that can get the answer right 80% of the time, and a smaller model that can handle smaller changes and get things right 60% of the time but correct its mistakes, then the system should try to run it on as many tasks as possible to save you money.. but in the end if ends up having to make a lot of corrections, then maybe you end up needing more total requests than the larger model. In that case maybe it's actually cheaper to run the larger model, if it takes fewer requests.

So I wonder how that kind of trade-off could be effectively calculated. I guess if you can figure out when "retries" happen you can count them and do some statistics on which model is more likely to work out in fewer shots. It's pretty complicated though, when you start to think about it in detail.

I do wonder if even having BOTH the smaller and bigger model make hypotheses, and try the smaller model's idea first, then if it fails, try the bigger model's idea, might be the way to go.

radarsat1··on Hacking Diffusion into Qwen3 for the Arc Challenge
Yes, I thought of the analogy with beam search but here I am proposing to add the backspace tokens to the context, not to actually rewind the context.
radarsat1··on Hacking Diffusion into Qwen3 for the Arc Challenge
Exactly what I had in mind. After generating a few wrong tokens perhaps the model could realize it leads to a low probability path and have a way to "go back", while staying in context. Parent is right though that thinking models can kind of do that without some special token, I hadn't thought about that, nice observation.
radarsat1··on Hacking Diffusion into Qwen3 for the Arc Challenge
Very interesting paper, even not considering the backspace stuff, thanks for the link. Pretty cool how that seems to tie in with more recent work on applying pure RL to LLM training.
radarsat1··on Hacking Diffusion into Qwen3 for the Arc Challenge
Regarding the typewriter approach, I've wondered for a while if anyone has explored simple backtracking with LLMs? Like, have the LLM be able to generate a backspace/delete token that lets it "undo" previously generated tokens in an append-only fashion. Not sure how this would work with teacher forcing but seems feasible with RL.
radarsat1··on Ask HN: How can ChatGPT serve 700M users when I can't run one GPT-4 locally?
I'm actually not sure I understand how MoE helps here. If you can route a single request to a specific subnetwork then yes, it saves compute for that request. But if you have a batch of 100 requests, unless they are all routed exactly the same, which feels unlikely, aren't you actually increasing the number of weights that need to be processed? (at least with respect to an individual request in the batch).
radarsat1··on Cursor CLI
Sounds like a job for a symlink
radarsat1··on Running GPT-OSS-120B at 500 tokens per second on Nvidia GPUs
Would love to try fully local agentic coding. Is it feasible yet? I have a laptop with a 3050 but that's not nearly enough VRAM, I guess. Still, would be interested to know what's possible today on reasonable consumer hardware.
radarsat1··on Running GPT-OSS-120B at 500 tokens per second on Nvidia GPUs
How long did your battery last?!
radarsat1··on Gemini CLI GitHub Actions
I'm honestly a bit confused by the free tier of Gemini. I've been using it with different agents (Aider, and then Crush), and I hit the rate limits FAST. Like, after maybe 5 or 6 requests it just blows up. Then I can try again quite a few times, and it hits the limit. Then eventually I guess I hit my daily limit and it just stops working until the next day.

I mean this has been enough to get my feet wet and have some fun with exploring agent-based development, no doubt, and I appreciate it, but I'm having a hard time crossing my experience with,

> generous free-of-charge quotas

as they say. It's not that generous if it stops working after 5 mins? (This morning literally a single sentence I typed into Crush resulted in some back and forth I guess it called the API a few times and it just rate limited-out. Fine, it was probably a lot of requests going on, but, but I literally gave it a single small job to do and it couldn't finish it.)

Meanwhile I seem to be able to use the Gemini web app endlessly and haven't hit any limits yet.

radarsat1··on Claude Code IDE integration for Emacs
There is ob-aider, maybe interesting for you to try.

https://github.com/emacsmirror/ob-aider

radarsat1··on Show HN: Jobs by Referral: Find jobs in your LinkedIn network
You mean you know.. most people?
radarsat1··on BorgBackup 2 has no server-side append-only anymore
I've been using btrbk with a local linux machine i use as a file server. Works well for incremental snapshot backups, no need to "unthaw" an image, I can directly fetch files from a previous snapshot. The only thing I haven't figured out with btrfs is how to efficiently handle incremental backs to S3. I guess there's not much choice than to use image diffs using btrfs-send because you don't have hard/ref links. But I don't like this because then if i want to retrieve a file from some version I'd have to have an extra 30 TB free to install the base image and progressively all the diffs up to the point I want to retrieve, seems a lot harder. So to make this reasonable I'd have to choose to make periodic non-incremental base images, starts getting complicated.
radarsat1··on Circuit Tracing: Revealing Computational Graphs in Language Models (Anthropic)
For a complex enough problem (like next word prediction on arbitrary text), I really have my doubts that any such method will result in an "interpretable" solution. More likely you end up with a giant stack of indecipherable if statements, gotos, and random multiplications. And that's assuming no matrices are involved, introduce those and you've just got a non-differentiable, non-parallelizable neural network.
radarsat1··on Lawrence of Arabia, Paul Atreides, and the roots of Frank Herbert's Dune (2021)
> I think he took things happening in small scale and used them as filler/detail to a story he wanted to tell in large scale.

I think this must be a common thing that good scifi authors do. I saw a talk by William Gibson many years ago where he mentioned something very similar, he called it "part of his science fiction toolkit".

radarsat1··on Scaling up test-time compute with latent reasoning: A recurrent depth approach
If you keep digging in that thread the author posts a gist containing information on how the recurrence works:

https://gist.github.com/JonasGeiping/65959599ca637d72d50c96c...

radarsat1··on Google is making AI in Gmail and Docs free, but raising the price of Workspace
I had the opposite experience recently. I was sent a summary of a sales video call, and the summary stated that we had promised to deliver something that was not nearly ready in 2 weeks! I was panicking but then started to doubt that the person in question would make such an irresponsible promise (but not.. completely sure it you know what I mean) so fortunately the summary included links to timestamps in the video call and I watched it. From the video it was clear he was talking completely hypothetically and not promising anything at all! The AI completely failed to pick up the nuance and almost made me change team priorities for the next sprint. Glad I verified it.
radarsat1··on Show HN: Value likelihoods for OpenAI structured output
and i guess includes other possibilities than numbers, like 'f' which could lead to four or five. There's probably a separate probability for 'fi' and 'fo' too.
radarsat1··on WorstFit: Unveiling Hidden Transformers in Windows ANSI
Seems like a another possible fix would be to change the best fit mapping table to never generate control characters, but only alphanumerics. So map quote-like characters to 'q' and so on.

This might be uglier and slightly change behaviour, but only for vulnerable applications.

radarsat1··on Maximum likelihood estimation and loss functions
Are you just demonstrating overfitting when estimating using too little data? Or is there something deeper going on in your example? What does the bank have to do with anything?
radarsat1··on OpenAI Reinforcement Fine-Tuning Research Program
Hi, the Zephyr link may be what I'm looking for. yeah I'm quite familiar with RL already so it was specifically RLHF that I was asking about, I'll check out that resource, thanks!
radarsat1··on OpenAI Reinforcement Fine-Tuning Research Program
If I'd like to learn more about DPO and RLHF, I've been looking for toy problems/datasets to use but coming up a bit empty handed. Is there a convenient way to experiment with these methods through toy problems and simulation that can be done on a single GPU? The need for massive data and parameter counts to do anything interesting makes learning about these methods a little daunting.
radarsat1··on ‘With brain preservation, nobody has to die’
In addition to all that we don't know about synapses etc, I've often wondered if even mapping all the "hardware connections" so to speak would even be enough. You'd have everything in the right place, but what about the "signals" running on it? Does a certain amount of constant activity on these circuits constitute signs of a "living" brain vs a dead one? How much of our consciousness is really in the topology of the circuits, and how much of it is simply defined by the constant activity running around in them? I assume neural circuits form loops that consist of synapses that reinforce or surpress activity. If these signals going around and around ever "stop", can they ever be started again with the same "patterns"? What if these patterns, the living "software", are at least partially what define you?

Well anyway that's my airchair crackpot neuroscience theory for the world to consume ;). I'm sure there must already be a name for the idea though.

radarsat1··on DeepThought-8B: A small, capable reasoning model
An algorithm that searches for the highest probability answer is not "reasoning"? "Search" has been a fundamental building block of GOFAI since the beginning. How do you define reasoning? Can you justify it being different from the last 70 years of thought on the topic?
radarsat1··on Why pipes sometimes get "stuck": buffering
Side note maybe, but is there an alternative to chaining two greps?
radarsat1··on Numpyro: Probabilistic programming with NumPy powered by Jax
Never heard of that, I'll take a look thanks. What do you mean by overly prohibitive? I'm not talking about big networks here so running them many times for inference is not a big deal, although I admit I'm not sure what numbers we are talking about. 100s to 1000s of times is feasible though.
radarsat1··on Numpyro: Probabilistic programming with NumPy powered by Jax
In some downstream applications such as filtering data (say, good/bad), I am training simple NN classifiers based on relatively small datasets. So, my personal confidence in the classifier is not so high, I'd like to reject things that are "definitely bad" and keep anything that may be good. Even more, I'd like to put aside "maybe good" data for human verification and keep "definitely good" data.

In other words, I think I have a practical use case for calibrated confidence scores, which I definitely don't get from my NN classifiers. They are right a certain percentage of the time, which is great, but when they are wrong sometimes they still have high confidence scores. So it's hard to make a firm decision based on the result without manually reviewing everything.

So my question is: is this an appropriate use case for PyRo? Will training my NN classifiers blindly converted to probabilistic classifiers and sampled appropriately give me actually reliable and useful confidence scores for this purpose? Is that the intended usage for this stuff?

radarsat1··on XMPP: The Gem of Instant Messaging
Not sure I buy that explanation. More likely is thst there's too little business advantage to being compatible with others and too much business advantage to having people inside your own ecosystem. Simple as that. There's just very little motivation for compatibility between these systems unless you are open source.
radarsat1··on Back to the future: Writing 6502 assembler with Amazon Q Developer
Imho these kind of pedantic comments detract from discussing the content of the article.
radarsat1··on State-space models can learn in-context by gradient descent
Converging slowly on Kalman filters, calling it now.
← PreviousPage 5 of 34Next →