popinman322

5 ms·

1.
by popinman322 2mo ago
With RAM prices how they are right now the used laptops might be more valuable than $1500.
2.
by popinman322 4mo ago
Very much agree. Until the vibe-coded version has been fully audited and profiled to perform, within reasonable tolerances, as well as the original code base, it feels like a bad idea to support it downstream or use it in production.
3.
by popinman322 4mo ago
Oh, this is great! I've filed bugs with JetBrains before and had them take months getting to my ticket, often with multiple hand-offs between team members; being able to provide a potential fix should make the process much faster.
4.
by popinman322 6mo ago
Does anyone know whether we'll be receiving transcoders for this batch of models? We got them for Gemma 3, but maybe that was a one-off.
5.
by popinman322 8mo ago
I've found that Gemini models often produce pseudocode that seems good at first glance but is typically wrong or incomplete, especially for larger or more complex functions. It might produce pseudocode for 70% of the function, then sil
6.
by popinman322 9mo ago
It doesn't look like the code anonymizes usernames when sending the thread for grading. This likely induces bias in the grades based on past/current prevailing opinions of certain users. It would be interesting to see the whole th
7.
by popinman322 10mo ago
They're comparing against open weights models that are roughly a month away from the frontier. Likely there's an implicit open-weights political stance here. There are also plenty of reasons not to use proprietary US models for co
8.
by popinman322 1y ago
You can swap experts in and out of VRAM, it just increases inference time substantially. Depending on the routing function you can figure out all the active experts ahead of the forward pass for a single token and pipeline the expert loadin
9.
by popinman322 2y ago
The executive branch is currently ignoring the law. Why would they start following it in 2029?
10.
by popinman322 2y ago
Not a fan of censorship here, but Chinese models are (subjectively) less propagandized than US models. If you ask US models about China, for instance, they'll tend towards the antagonistic perspective favored by US media. Chinese model
11.
by popinman322 2y ago
Assuming you're doing local inference, have you tried setting a token filter on the model?
12.
by popinman322 2y ago
DeepSeek was built on the foundations of public research, a major part of which is the Llama family of models. Prior to Llama open weights LLMs were considerably less performant; without Llama we might not have gotten Mistral, Qwen, or Deep
13.
by popinman322 2y ago
It's always very interesting to see people pull out threads with low like counts (like 12k) and claim that central idea of the post is widely held. We're talking about platforms with tens of millions of users; wide appeal is at le
14.
by popinman322 2y ago
The LSP is limited in scope and doesn't provide access to things like the AST (which can vary by language). If you want to navigate by symbols, that can be done. If you want to know whether a given import is valid, to verify LLM output
15.
by popinman322 2y ago
Google Trends make it seem like we're out of the exponential growth phase for LLMs-- search interest is possibly plateauing. A decline in search interest outside of academia makes sense. The groups who can get by on APIs don't car
16.
by popinman322 2y ago
Try LiteLLM; their core LLM proxy is open source. As an added bonus it also supports other major providers.
17.
by popinman322 2y ago
Huge +1. If I'd understood this mantra earlier in my career it would have saved me a large amount of hassle. For juniors: any time you send something important to your manager, confirm they read the document. Don't ask "did y
18.
by popinman322 2y ago
Grep is also useful when IDE indexing isn't feasible for the entire project. At past employers I worked in monorepos where the sheer size of the index caused multiple seconds of delay in intellisense and UI stuttering; our devex team&#
19.
by popinman322 2y ago
It's possible to get stuck in merge hell where all your reviewers ok the PR but someone merged a conflict 2 seconds ago, or you've got a reviewer in Singapore while you're in SF and conflicts appeared overnight. In general it
20.
by popinman322 2y ago
This is where supporting machinery & RAG are very useful. You can auto- lint and test code before you set eyes on it, then re-run the prompt with either more context or an altered prompt. With local models there are options like steerin
21.
by popinman322 2y ago
Tangent here: really? I've found base Whisper has concerning error rates for non-US English accents; I imagine the same is true for other languages with a large regional mode to the source dataset. Whisper + an LLM can recover some of
22.
by popinman322 2y ago
Also, similar to Orca-Math but without a teacher model. They also followed an iterative DPO/KTO scheme, but with no length normalized NLL loss term.
23.
by popinman322 2y ago
I remember stumbling upon an early discussion about this [0] a bit ago in the EleutherAI discord when searching for discussion about a paper; I'm glad to see it's turned into something public. [0]: https://discord.com&#
24.
by popinman322 2y ago
Any news on when Unsloth's parallel full tuning will be available?
25.
by popinman322 3y ago
You could likely also combine the LLM with a coverage tool to provide additional guidance when regenerating the fuzzer: "Your fuzzer missed lines XX-YY in the code. Explain why you think the fuzzer missed those lines, describe inputs t
26.
by popinman322 3y ago
vs RAG: RAG is good for searching across >billions of tokens and providing up-to-date information to a static model. Even with huge context lengths it's a good idea to submit high quality inputs to prevent the model from going off o
27.
by popinman322 3y ago
For the tasks my group is considering, even a 7B model is adequate. Sufficiently accurate responses can be fed into other systems downstream and cleaned up. Even code responses can benefit from this by restricting output tokens using the gr
28.
by popinman322 3y ago
Previous discussion, on a link to the implementation: https://news.ycombinator.com/item?id=37740932
29.
by popinman322 3y ago
I don't have my phone number attached to my GitHub account, if that's any indication
30.
by popinman322 3y ago
As someone who participates in relatively high information density Discord servers (mostly ML related), search is rarely a problem for me. Yes, you need to do keyword search, but it's not the end of the world-- whenever I've neede

More ›