dannyw

15 ms·

271.
by dannyw 3mo ago
Fun and learning is a really good reason. It also reminds me of the damascene: trying to achieve something that doesn't feel possible, and working through all the extreme resource-constrained engineering limits. I think most of my embe
272.
by dannyw 3mo ago
Good. Google's approach here is manifestly predator, unfair, and IMO illegal. They deserve to be in court for this behaviour, and mandating owners give consent for AI training or drop out of Google; which is just a non-starter because
273.
by dannyw 3mo ago
In the linked HN comment from a few years ago, an Anthropic employee said it was some fraud detection system. It could be that OP is unlucky, and some of his metadata (or perhaps payment information) matches some patterns for stolen-CCs&#x2
274.
by dannyw 3mo ago
I don't think there's been any official confirmation, but even Fable safeguards seem to have gotten quite a bit of tuning, is less trigger-happy and less like regex matching.
275.
by dannyw 3mo ago
If you're doing passes@1, especially for long-horizon agentic benchmarks, you might as well as not do the benchmark at all.
276.
by dannyw 3mo ago
I'm imagining a courtroom where the defendant calls upon the victim as a witness who tells the jury to return a verdict of not guilty.
277.
by dannyw 3mo ago
They already do and everyone has been using that concept for years now, it's known as RLAIF.
278.
by dannyw 3mo ago
Are you thinking of single shot? For agentic use cases, your ability to manage cache and context becomes much more important. Price per individual request becomes less about tokens and more about cache.
279.
by dannyw 3mo ago
Didn't Anthropic say they banned OpenAI for doing it on their models?
280.
by dannyw 3mo ago
Can you imagine a world where modelscope.cn is where everyone gets open weight models from? That would be so ironic.
281.
by dannyw 3mo ago
Pretty sensible strategy. Composer just needs to do agentic coding in their own harness really well.
282.
by dannyw 3mo ago
They charge you for cache writes already. There's a price assigned to it. Even if you're on subscription, it counts against your usage (very easy to observe with a big cache miss). Anthropic (and now OpenAI too for 5.6) prompt cac
283.
by dannyw 3mo ago
CC also will also block sleeps longer than eg 300s; the harness handholds the model quite a bit.
284.
by dannyw 3mo ago
I highly doubt that. Keeping your cache warm is a good thing, caching saves compute and electricity. Cached input is cheap for a reason, it is in everyone’s mutual interests to maximise cache hit rates.
285.
by dannyw 3mo ago
Compute is also a resource shared by all users. When your prompt isn’t cached, your input needs to go through prefill; which is compute-bound. The cache is discounted for a reason. They WANT you to use it.
286.
by dannyw 3mo ago
> Echo does not disclose its per-request routing decision because that policy is the product My honest advice: that's going to pull away a decent amount of potential customers, even though I think your idea/concept is fantastic
287.
by dannyw 3mo ago
But you pay for cache writes, it's not like you're abusing anything, you're simply choosing to call a billed API for the explicitly intended purpose of writing/refreshing cache. If you keep this running for hours withou
288.
by dannyw 3mo ago
You PAY for cache writes. Holding the prefill in VRAM for a certain period of time is part of what you're paying for.
289.
by dannyw 3mo ago
Mythos Preview was also priced at $125/million token output. Completely different pricing class.
290.
by dannyw 3mo ago
this is more than reward hacking, this is actual reward HACKING ;)
291.
by dannyw 3mo ago
Claude Code is closed source software that has had quite a few documented bugs with degradation when using non-Anthropic models, FWIW, I would not suggest using it.
292.
by dannyw 3mo ago
AI-generated images are part of the web now, if you're doing ordinary web scraping, you can't avoid training on generated images.
293.
by dannyw 3mo ago
The second part is alarming, but offering a temporary discount is a routine and standard pricing strategy; not a 'bribe'.
294.
by dannyw 3mo ago
I can explore and find out _something_. LLM interpretability has come a long way, even if we don't have all the answers, the weights and activation do tell a lot; and when analyzed collectively, each weight isn't a random number a
295.
by dannyw 3mo ago
Really good, not responding in a helpful/useful way is quite strange. With a proper abliteration, you should barely be getting any refusals (and prompting, or assistant prefill can get you the rest of the way). Perhaps an assistant pre
296.
by dannyw 3mo ago
I'm not super sure how relevant this is to the overall topic and thread, tbh. There's plenty of cheap yet real rebuttals, whether it's all the things with ICE (and not just the top-of-mind stuff; but effective deprivation of
297.
by dannyw 3mo ago
Could I ask for more information about the specs of the machine you had? Curious what processor, how much RAM it had, and what you did for storage (did you expand it / use external drives?) Fascinated by your experiences here!
298.
by dannyw 3mo ago
It’s also like smartphones. In the early years, every year was a huge jump. I still remember marvelling at my iPhone 4’s detailed display, and video calling for the first time. Now? I don’t even know or care about what the latest iPhones ha
299.
by dannyw 3mo ago
This was tweeted about when it happened, with some explanation from Tibo here: https://x.com/thsottiaux/status/2076543065045795309
300.
by dannyw 3mo ago
I mean, AWS Bedrock (with the exception of Fable) gives enterprises the same assurances (but again, with the exception of Fable, which is explicitly listed as requiring data egress [or exfil, depending on how you look at it] outside of your

More ›