nl

15 ms·

361.
by nl 4mo ago
The design on the system is specifically done to make that impossible because you control the encryption keys and they can't see your data. It nearly works except for the annoying hardware signing keys. With OHTTP it might still be
362.
by nl 4mo ago
> Anthropic confirmed the training run was under 10^26 flops. You can't train 10T to chincilla and stay under 10^26. I don't think Anthropic have said anything of the sort. Microsoft published it as 6.1*10^27 FLOPs[1] Elon has
363.
by nl 4mo ago
> Rented out GPUs are likely not a similar use profile as compute used for training LLMs. The latter is likely closer to the cryptocurrency GPUs that are running at full tilt 24/7. This is untrue. H100's are used for training (
364.
by nl 4mo ago
Yes, it's entirely possible to do that via RL. It'd be a fun little project you could do for less than $100 on a small LLM actually.
365.
by nl 4mo ago
The argument is that all GPUs are currently appreciating (!!) Google is still running 10 year old Tesla T4s at full capacity. This is way beyond the expected lifetime.
366.
by nl 4mo ago
Memory is unlikely to drop in price before mid-2027 when new capacity starts to come online. The GPU shortage looks to be even longer lived.
367.
by nl 4mo ago
For practical purposes llama.cpp is this. You can link to it or use the network API.
368.
by nl 4mo ago
> Compute is also a rapidly depreciating asset. That's the default assumption but in the new GPU+Memory constrained age isn't true. Time on 4 year old H100 servers costs more now than when they were new (!!)
369.
by nl 4mo ago
Their models are much smaller: 1T vs 5T for the frontier models. 1T is Sonnet/Google Flash size, not Opus size. The $0.87/M tokens price for Mimo Pro is probably subsidized. Mimo models aren't widely available on western prov
370.
by nl 4mo ago
> No one is bitter lesson pilled anymore. Will the 10T parameter Mythos model be released this month or next month? They better soon because it is generally accepted that one of the reasons GPT 5.5 is better at hard tasks than Opus is be
371.
by nl 4mo ago
Actually in this case they possibly are estimates. It's been known for some years[1] that LLMs do regression in-context. Frontier models have been trained against many, many issue text that include task break downs and estimates. [
372.
by nl 4mo ago
Apple's PCC is the best option for this kind of offload that exists. However the PCC root keys are still signed by Apple which requires you to trust Apple and the laws in the jurisdiction Apple operates in . Edit: for this update they
373.
by nl 4mo ago
Keeping your account secure is not the same as maintaining a project. But it is a great example of the social contract! If you fail to keep your account secure you lose your account. If you don't maintain your project then someone fork
374.
by nl 4mo ago
You understand this is the opposite of your example before right? It's like a business asking for volunteers, you saying you will, then the business demanding that you turn up when it suits them and you not being allowed to say &quot
375.
by nl 4mo ago
This just isn't true. Reliance requires an exchange of value. If you get something for zero value without a contract it's a gift under US commercial law. You need to provide citations if you keep insisting otherwise because every
376.
by nl 4mo ago
In the long term all is dust. NVIDIA has at least 2 years of solid revenue growth ahead of it. Beyond that people are dreaming about doing predictions anyway.
377.
by nl 4mo ago
Fortunately this is completely wrong.
378.
by nl 4mo ago
It might pay you to look at the numbers. Last quarter: Goog: $109B revenue, $62B in profit. NVIDIA: $81B revenue, $58B in profit. NVIDIA is growing faster. I agree about Telsa though.
379.
by nl 4mo ago
> Selling a toaster has an implicit warranty of merchantability. Society expects that if you sell me something, it should have certain promises. Yes, there’s no monetary exchange here, the work is given gratis, but there’s still a relati
380.
by nl 4mo ago
> However, I strongly believe that when you create something for people to use that there’s an implicit social contract about how to go about doing certain things. Wow. The entitlement in this statement is outrageous.
381.
by nl 4mo ago
Qwen 3.5 35B A3 Qwen models are always good. The 35B A3 model is a MoE model which means it has higher performance in RAM constrained environments compared to the 27B dense model (which is better at coding). I don't have experience to
382.
by nl 5mo ago
I believe democratisation of production and distribution is a net good thing. You don't. That's not about technology but is a fundamental moral issue.
383.
by nl 5mo ago
This has been the case for longer than I've been alive. Have you not seen the old Soviet photos where people who fell out of favour were erased from photos? Media literacy has always been about working out if you can trust something. N
384.
by nl 5mo ago
I LOVE being able to create images of things I could previously only imagine. I think the inability to see the freedom AI gives people is one of the saddest things I've seen. I remember when the internet was young people would compla
385.
by nl 5mo ago
> Current inference providers are being subsidized with VC money at this point. This isn't true. Anthropic is making an operating profit including the loss making subsidised subscriptions (but excluding training). Your normal infe
386.
by nl 5mo ago
This wasn't possible 12 months ago, and barely 6 months ago.
387.
by nl 5mo ago
Anthropic explicitly state that they don't do this, even if you use the free plan and even if you don't opt-out of letting them use your data for training: "We do not sell users’ data to third parties." https:/&#x2
388.
by nl 5mo ago
> Plus, there are no RLHF signals in OpenRouter data. Even if OpenRouter wanted to build a general model-neutral framework for collecting RLHF-type data, it can't force subscriber apps to do the UI-level stuff necessary to collect i
389.
by nl 5mo ago
> You should expect that any inputs and outputs are going into someone's training database. OpenRouter explicitly lets you filter by zero-data-retention providers: https://openrouter.ai/models?zdr=true
390.
by nl 5mo ago
This isn't what parallel reconstruction means. This seems to be reverse engineering the attack.

More ›