dannyw

16 ms·

391.
by dannyw 4mo ago
When a large open weight model is released, a lab, startup, or a rich hoist can easily do logit-level distillation and create a XXb param model or whatever, and in theory you should get a really good distill.
392.
by dannyw 4mo ago
DGX Spark is ~273GB/s. That’s about M5 Pro territory, and twice as fast as the M5. You’d have to go to the M5 Max, or M3 Ultra, to get higher memory bandwidth than the Spark.
393.
by dannyw 4mo ago
The DGX Spark has basically the same memory bandwidth as a M5 Pro, and far more than a M5. Only the M3 Ultra really beats it, and once you start scoping out the cost of a M3 Ultra with 128GB or 256GB, the DGX Spark doesn’t look bad after al
394.
by dannyw 4mo ago
It’s less about 1 scalper buying 100. Order limits, credit card limits, etc already restrict that. It’s more about 1000 scalpers buying two or three to immediately resell. There are entire discords and communities around this with thousands
395.
by dannyw 4mo ago
Interestingly not Opus 4.5 apparently, which is still available via API.
396.
by dannyw 4mo ago
Yes, OpenAI admits they silently reroute sensitive requests to different models for user welfare at least: https://openai.com/index/building-more-helpful-chatgpt-exper... The implementation was so borked, SamA went bac
397.
by dannyw 4mo ago
Also their API customers and downstream customers (e.g. Cursor users) would also need similar infra, and probably a decent amount of users would just choose another model that doesn't require ID & an immigration status check. And A
398.
by dannyw 4mo ago
The Internet will find a way to route around censorship. I’ve already started downloading notable open weight models onto my NAS and will distribute via eg BitTorrents if needed. I’m not in the US, and you can’t ban VPNs overnight.
399.
by dannyw 4mo ago
Maybe, but the world and the internet isn’t just the US. Businesses outside of the US, like the EU, might have significant competitive advantages.
400.
by dannyw 4mo ago
They are open weight, so you can abliterate: https://github.com/p-e-w/heretic You can finetune and mould it to whatever you want.
401.
by dannyw 4mo ago
Let LLMs help you with coding. Design the game and the mechanics yourself. I can see this being an incredibly empowering tool in the right game developer's hands; but if you come into it with a token-maxxing / AI-maxxing mentality
402.
by dannyw 4mo ago
I feel really bad for Anthropic right now. This should never have happened and seems like another arbitrary use of government power, Friday after market closes. Whatever you feel about Anthropic, good or bad, this is not fair, and this is n
403.
by dannyw 4mo ago
It seems like this might not have happened if Anthropic didn't institute the cyber blocks as broadly to also cover pure-defensive use cases? Because this wouldn't be considered a jailbreak with any other model; which would just do
404.
by dannyw 4mo ago
Yeah, I'm really not sure what the point of this paper is. Every non-toy environment does prefix caching.
405.
by dannyw 4mo ago
In particular, mental health.
406.
by dannyw 4mo ago
I just find doubtful that a LLM is going to help, instead of hurt, any state actor that is capable of starting a nuclear weapons problem.
407.
by dannyw 4mo ago
I meant a strict budget given by my parents (and I could ask for more with justification). One of the valuable lessons I have learned is that there's no spending caps on AWS, but it taught me to set up billing alerts :)
408.
by dannyw 4mo ago
AWS will likely write off most costs automatically, but if you truly do manage to rack up a $50k bill somehow, you're getting sent to collections and/or their legal team. The terms you signed obligate you to pay your balance. Whet
409.
by dannyw 4mo ago
Honestly, probably not that much electricity. AWS will charge you the hourly price irrespective of your load/power consumption. But instances sitting idle generally don't use that much power.
410.
by dannyw 4mo ago
Most modern LLMs (especially frontier ones) are large token hogs because they draft, check, re-draft, the content (whether an output message; or a code diff) sometimes multiple times in the thinking block. When you see a thinking summary li
411.
by dannyw 4mo ago
Obviously the specifics vary by jurisdiction, but usually contracts that are 'necessary' (e.g. grocery store purchases) or beneficial to the minor (e.g. an employment agreement) cannot be voided simply because someone is under 18.
412.
by dannyw 4mo ago
I couldn't disagree more. I was playing around with AWS when I was probably 14 years old, with a credit card from my parents with consent, and a strict budget and the understanding that if I mess up and overspend, I'm getting disc
413.
by dannyw 4mo ago
I burst out laughing when the agent spawned a subagent to join IRC. So funny.
414.
by dannyw 4mo ago
Yeah, I think Fable is really good for debugging tricky bugs. Setting boundaries in your prompt / markdowns helps; for example if I tell it to not use any web browser automation, I have seen Fable respect both the rule and the spirit o
415.
by dannyw 4mo ago
No, a company choosing to use some terminology doesn’t make it correct nor canonical in any sense; especially when they have a vested interest in not being neutral or credible. If Google starts calling ads “Best Links” that doesn’t make it
416.
by dannyw 4mo ago
The Anthropic refusal description is even more direct. “The request could assist the development of competing AI models, which is restricted under Anthropic's commercial terms. Benign machine learning work can also trigger this categor
417.
by dannyw 4mo ago
This is the beauty of open source :) KHTML -> WebKit -> Blink is a good example.
418.
by dannyw 4mo ago
MiMo Code is not a model, it's a harness like Claude Code / OpenCode / Codex (which is still open source, Apache 2.0, btw). You might mean the MiMo-V2.5-Pro model?
419.
by dannyw 4mo ago
I use Codex and wanted it to sort through the footage and use subagents to review. Codex limits are fairly generous, esp paired with mini models for this kind of task generally, but even GPT5.5 usage is still pretty generous. Again, it’s th
420.
by dannyw 4mo ago
Some anecdotal social reports seem to suggest it wasn’t just giving suboptimal answers, but rather mucking around and sabotaging your codebase and training (like editing hyperparameters in project files despite not being requested). Of cour

More ›