BoorishBears

5 ms·

61.
by BoorishBears 3mo ago
If they were on open weights, at some point the provider would deprecate it, probably with worse notice And self-hosting would probably have been more expensive unless they had massive volume: Deepseek V3.x would have been the comparable op
62.
by BoorishBears 3mo ago
No one's firing up a residential proxy to read your blog, and the corporate AI scrapers have all the resources in the world even without residential proxies They're most useful for getting information from the cloud hosted sites t
63.
by BoorishBears 3mo ago
"Raises the question of what we see is real" No they really don't, dishonest founders do that. You're one with the lower case shibboleth so I have no doubt you surround yourself with dishonest founders, but faking users
64.
by BoorishBears 3mo ago
Are you joking or are you confusing huge valuations with huge headcount?
65.
by BoorishBears 3mo ago
Thank god for residential proxies. Highly unethical but the way the internet is going they're the last anti-hero of a somewhat open internet
66.
by BoorishBears 3mo ago
This would feel less gross if Boom would do it themselves to show they know how to. If even Boom can't do it, then this prize is cartoonishly low. They're offering 750k in cash and a Toyota Highlander in Boom stock to a team that
67.
by BoorishBears 3mo ago
The first thing I read from you was a sardonic browbeating in response to the exact comment I gave an earnest response. And even in domains that lean heavily on "usual phrasing", like technical writing, human writing has notably h
68.
by BoorishBears 3mo ago
Did you try reading the whole comment? Once creativity is being measured in isolation, getting multiple responses from the model is enough to measure creativity a ton of different ways: wordfreq to identify overused phrases, getting multipl
69.
by BoorishBears 3mo ago
The largest model I've post-trained in the last 2 years of working on this problem was Kimi 2.5 at 1T parameters. The simplest way I'd put it is, teaching a model to write coherently (follow rules, patterns, etc.) is easy enough:
70.
by BoorishBears 3mo ago
I'm pretty sure Altman has spoken about giving a model 100k+ A100s specifically, this might be them being very literal
71.
by BoorishBears 3mo ago
"It'd be a game-changer to be able to have the model start replying with partial information streamed from the tool call, then seamlessly continue with additional information." You can't continue a conversation if it hin
72.
by BoorishBears 3mo ago
I used to pride myself on not being the "fonts too pointy, scroll too buttery" crowd! But AI has brought me full circle and now nothing removes my interest in reading even a single word on a page faster than purple gradient greebl
73.
by BoorishBears 3mo ago
> based on FP4 FLOPs (yes it's disadvantageous to A100, that's the point) The A100 doesn't have hardware FP4, and you'd be running a quantized model with some accuracy loss but unless this was natively trained on FP4*
74.
by BoorishBears 3mo ago
Writes pretty engaging prose, finetunes well, now MIT licensed... what's not to like? Oh and very good world knowledge for the size: better than than DS4 Flash
75.
by BoorishBears 3mo ago
It rubs me the wrong way that someone would respond to something I need with 0 context just to tell me it's not a big deal. Some people have standards in what they build.
76.
by BoorishBears 3mo ago
Do integrations supporting streaming input? One big gap I've run into for UX is most realtime voice harnesses wait for a full response from tools, and at most support the model filling the dead air until then It'd be a game-change
77.
by BoorishBears 3mo ago
We came full circle! https://news.ycombinator.com/item?id=45393382 https://news.ycombinator.com/item?id=45398577 I've said before: > Vercel is not a problem. Pumping millions of dollars into the JS
78.
by BoorishBears 3mo ago
Both of you are right of course
79.
by BoorishBears 3mo ago
That's not a real problem, hardly any 3rd party was running native weights anyways so they'll get quantized to NVFP4
80.
by BoorishBears 3mo ago
The only thing that diminishes the value of a GPU right now is unsupported features with outsized value during inference and/or training (like FP4 support) and it takes time for those features to actually take off And labs are fully le
81.
by BoorishBears 3mo ago
> in 18-24 months when they cost significantly less ... going to need you to sit down for this one...
82.
by BoorishBears 4mo ago
Moving forward requires two parties with two very different financial incentives to cooperate in a way that harms the incentive of the other: the manufactures making the chips, and the companies using the chips to build AI. Even if the Chin
83.
by BoorishBears 4mo ago
> After July 7, 2026, Claude Fable 5 is no longer included in your plan’s weekly usage limits. You can keep using Claude Fable 5 through usage credits, which let you pay for usage beyond what your plan includes. They stated the desire n
84.
by BoorishBears 4mo ago
> After July 7, 2026, Claude Fable 5 is no longer included in your plan’s weekly usage limits. You can keep using Claude Fable 5 through usage credits, which let you pay for usage beyond what your plan includes. This is the real story.
85.
by BoorishBears 4mo ago
They're referencing Gemini 3.5 Flash being the top model, you must not be great with detail. And no (strong) programmer would jump to assuming other people are coding monkeys just because they disagree on what a strong LLM is: that&#
86.
by BoorishBears 4mo ago
There are image models with censorship at every stage from pretraining to posttraining. Most recently Ideogram released an open weight model that will denoise into a grey image with the text "Blocked by safety filter" notice for c
87.
by BoorishBears 4mo ago
Karma systems are never perfect, and most people will not assume this is a pattern. (ie. won't feel the need to downvote them just for having yet another crappy AI benchmark) I only recognize it because I build a product that leaves me
88.
by BoorishBears 4mo ago
This guy had a terrible broken benchmark that gets hawked every release, and I wish HN would ban accounts that essentially exist to hawk a personally owned site, especially such a bad one.
89.
by BoorishBears 4mo ago
Specifically Dwarkesh couldn't understand that GPUs are not enough: it's GPUs plus multiple ecosystems to leverage them at massive scale during training vs inference. Instead of giving China open access to US controlled chips and
90.
by BoorishBears 4mo ago
If they really managed this from pre-training a 1.6 T parameter model through to post-training without NVIDIA, Dwarkesh Patel got what he wanted.

More ›