jsnell

16 ms·

241.
by jsnell 1y ago
Fair enough! But that's not real-time communication during an active incident. It's communication O(days) later.
242.
by jsnell 1y ago
The status page is not where you communicate about either the root cause or about the action plan for preventing recurrences.
243.
by jsnell 1y ago
> Protobuf's original sin was failing to distinguish zero/false from undefined/unset/nil. It's only proto3 that doesn't distinguish between zero and unset by default. Both the earlier and later versions supp
244.
by jsnell 1y ago
Discussed many times over the years: https://news.ycombinator.com/item?id=18188519 (299 comments) https://news.ycombinator.com/item?id=21871514 (215 comments) https://news.ycombinator.com/it
245.
by jsnell 1y ago
Apple has been shipping device attestation in their browser for years (Private Access Tokens), with no backlash.
246.
by jsnell 1y ago
> OpenAI is making $12 billion a year off of 700 million users [1], or around $17 per user annually. What other products that have no ad support perform that badly? That's a bit of a strange spin. Their ARPU is low because they are
247.
by jsnell 1y ago
It's a business with odd scaling. You can find examples of national papers of record having successful subscription models for text content. If you're only subscribing to one paper, it'll be one of those. And you can find 1-5
248.
by jsnell 1y ago
Clearly that's not the use you have for a new XML library. It's a use you're imagining somebody else would have for it. And because you're just imagining the use case, you've failed to think through what "goo
249.
by jsnell 1y ago
I suspect somebody could. But why? Do you need a new XML library? I don't. And if you don't need one, why write one? If there is no specific use case in mind, how do you even determine what dimension "good" is measured o
250.
by jsnell 1y ago
If it were a lie, why take the PR hit of telling the truth about starting to train on user data but lying about the specifics? It'd be much simpler to just lie about not training on user data at all. If your threat model is to uncondit
251.
by jsnell 1y ago
Billions of people use Google, YouTube, Facebook, Tiktok, Instagram, etc and accept the ads. Getting similar ad rates would make OpenAI fabulously profitable. They have no need to start with ad formats that might be rejected by users. Eve
252.
by jsnell 1y ago
The capital costs for the GPU are an order of magnitude larger than the energy consumption. It doesn't matter whether the GPUs are used for training or inference. Back of the envelope: $25k GPU amortized over 5 years is $5k/year.
253.
by jsnell 1y ago
An unavoidable aspect of abuse problems is that there is no perfect solution. As the defender, you’re always making a precision vs. recall tradeoff. After you’ve picked off the low hanging fruit, most of the time the only way to increase re
254.
by jsnell 1y ago
> why aren't there multiple API providers offering models at dirt cheap prices? There are. Basically every provider's R1 prices are cheaper than estimated by this article. https://artificialanalysis.ai/models&#x
255.
by jsnell 1y ago
This has no connection with reality. This is not an attestation mechanism, and can't be used as one.
256.
by jsnell 1y ago
I linked to the writeup by Deepseek with their actual numbers from production, and you want "better numbers" than that?! > Each H800 node delivers an average throughput of ~73.7k tokens/s input (including cache hits) durin
257.
by jsnell 1y ago
For the top few providers, the training is getting amortized over absurd amount of inference. E.g. Google recently mentioned that they processed 980T tokens over all surfaces in June 2025. The leaked OpenAI financial projections for 2024 sh
258.
by jsnell 1y ago
I don't believe the asymmetry between prefill and decode is that large. If it were, it would make no sense for most of the providers to have separate pricing for prefill with cache hits vs. without. Given the analysis is based on R1, D
259.
by jsnell 1y ago
Paywalled.
260.
by jsnell 1y ago
Why make up such an absurd grievance to get fake-outraged about? Nowhere in the article does it say that Google claims to have invented liquid cooling. In fact, nowhere does it say they claim any part of this is a new invention. But the poi
261.
by jsnell 1y ago
Text slides interspersed with the transcript work great normally. But this is mostly slides with just four words of content, and a transcript that just be repeats the same words.
262.
by jsnell 1y ago
> GPT-5, Claude, and Gemini represent remarkable achievements, but they’re hitting asymptotes This part could do with sourcing. I think it seems clearly untrue. We only have three types of benchmark: a) ones that have been saturated, b)
263.
by jsnell 1y ago
I know there's a lot of rebuttals to this statement already, but I think there's a simpler way of showing it is incorrect: Figure 2 in the paper shows the LMArena score of whatever model is used for "median" Gemini query
264.
by jsnell 1y ago
The report also doesn't define what the word "AI" means. What are they trying to hide?! In reality, we know what Google means by the term "Gemini Apps", because it's a term they've had to define for e.g. t
265.
by jsnell 1y ago
No space between paragraphs and blockquotes definitely hurts, and the subheading formatting is clearly not right. But the same content is posted on multiple sites. Maybe you'd find e.g. the Substack formatting more readable? https:&#x
266.
by jsnell 1y ago
Yes, but it's not going to be 0.02 USD of PoW per page! That is an absurd number. It'd mean a two-hour proof of work for a server CPU, a ten hour proof of work for a phone. In reality you can do maybe a 1/10000th of that befo
267.
by jsnell 1y ago
GPT-4 was a long time ago, and honestly mostly useless. But a lot of that progress was already present in the intervening models, and it's easy to forget it happened when comparing GPT-5 to the state of the art a month ago rather than
268.
by jsnell 1y ago
The entire problem is that proof of work does not increase the cost of scraping by 100x. It does not even increase it by 100%. If you run the numbers, a reasonable estimate is that it increases the cost by maybe 0.1%. It is pure snakeoil.
269.
by jsnell 1y ago
A proof of work challenge does not throttle the scrapers at steady state. All it does is add latency and cost to the first request.
270.
by jsnell 1y ago
The human needs to wait for their computer to solve the challenge. You are trading something dirt-cheap (CPU time) for something incredibly expensive (human latency). Case in point: > If the challenge takes, say, 250 ms on the absolute b

More ›