wolttam

4 ms·

31.
by wolttam 2mo ago
I would have thought the Amazon tax is “free shipping” that has altered the pricing of every item, and now B&M retail is picking up the same pricing despite a much lower shipping overhead Ads too, I guess
32.
by wolttam 2mo ago
+ issues + pull requests. There’s some nice to haves in there, that you’d need some other service or convention to manage
33.
by wolttam 2mo ago
I’m getting good at reading LLM-speak (which is probably sad) It means two “layers” of information access/validation. Sounds like it first made a little map of model <-> benchmark, then went and filled in the score boxes. Definit
34.
by wolttam 2mo ago
I think an argument can be made that some software has gotten more secure. I don’t think it can be denied that the models do show an ability to find security vulnerabilities that may have otherwise been missed
35.
by wolttam 2mo ago
It feels icky to me that they feed a log of what you’ve watched, when, and on what device, over to an LLM just to get back some recommendations. Even though It’s probably exactly the type of thing I would do on a locally hosted LLM
36.
by wolttam 2mo ago
So you’d be attaching a new volume to the running worker VM for each PVC? This seems a little odd to me. Could you attach a single large volume, and do path-based provisioning on that?
37.
by wolttam 2mo ago
Opus 4.5, even 4.6-level performance has been around since July 31st in 284B total params and just 160GB of weights at native FP4 quantization- DSv4 Flash.
38.
by wolttam 2mo ago
I think I saw a better overall composition out of Flash 0731 Effort on this one?
39.
by wolttam 2mo ago
See also FreeTube
40.
by wolttam 2mo ago
Live in an overall less uncertain environment?
41.
by wolttam 2mo ago
My comment was trying to point on the inanity of calling out "so many em-dashes" on a post where it doesn't seem that obvious it was LLM generated. I added some of my own inanity with my remark about them being double-dashes,
42.
by wolttam 2mo ago
There really wasn’t that many and this did not have the voice of written-by-an-LLM to me And they’re not em-dashes -- they’re double hyphens.
43.
by wolttam 2mo ago
I think you’re thinking about it in slightly the wrong way. We’re not throwing more data at frontier models in hopes they get more/better capabilities somehow. We’re either: setting up a verifiable task, and doing RLVR to get the model
44.
by wolttam 2mo ago
They make money on each token when you look at the electricity and interconnect fees, but no, I don’t think they’ve turned a profit on their Capex, even a little bit
45.
by wolttam 2mo ago
What areas do you think model capability will plateau in, and why?
46.
by wolttam 2mo ago
As written it sounds like you're saying that it was incompetent of the labs to disable the sandbox internet access? They tried to disable open internet access but the models zero-day'd their Artifactory package registry and got in
47.
by wolttam 2mo ago
Automated defence is going to use so many tokens.
48.
by wolttam 2mo ago
Yours is the only mention of GPT-OSS in the whole thread. I count about 2-3 comments saying the model is meh and many more saying it’s a big step up. I am among those with real life experience with the model that used the previous as well a
49.
by wolttam 2mo ago
I think we will eventually reach a point where this is the case, but at the moment it seems like you can throw virtually any non-trivial use-case at a model today and end up being more satisfied with the results that a model tomorrow gives.
50.
by wolttam 2mo ago
It feels like the rate of stupid we have achieved in the last (checks watch) ~18 months has surely eclipsed anything that may have occurred during the roman empire. By sheer volume alone?
51.
by wolttam 2mo ago
It's a terrible moat. You etch the silicon then nobody wants to run it in 6 months because models have advanced that much further.
52.
by wolttam 2mo ago
We’re already seeing incredible knowledge compression out of the models the GP mentioned like Qwen 35B-A3B, which feels well within the realm of “runs on a phone” in the next handful of years. And by then we’ll probably have been further su
53.
by wolttam 2mo ago
Essentially, MCP is the harness’ wheelhouse, not the model’s
54.
by wolttam 2mo ago
MCP is the protocol by which the harness can connect with additional tools. The harness does the work of discovering the tools, the model is then fed a description of what tools it has available, with zero awareness of whether those tools a
55.
by wolttam 2mo ago
MCP is invisible to the LLM, or should be, under normal use
56.
by wolttam 2mo ago
The problem is that there is no signal to the folks involved whether you have bothered to understand what you’re posting or are simply offloading that work to others. Same issue as dumping a system log or a stack trace without any of your
57.
by wolttam 2mo ago
2 sparks currently run this model at 60 t/s single session, up to just over 100 t/s aggregate with concurrency of 4. Going local has as opened up a world of use-cases I never would have entertained the idea of on metered/clo
58.
by wolttam 2mo ago
Hooray! This model makes me very optimistic about the future of local inference. The CyberGym score stands out to me.
59.
by wolttam 2mo ago
It runs really well on 2 DGX Sparks - 60t/s
60.
by wolttam 2mo ago
> and used tiktoken to approximate tokens, by dividing character count by four. This is just the type of thing that stands out. You used a tokenizer library to determine the number of tokens in your text by dividing its length by 4. That

More ›