Maybe recruited by saudi intelligence. The AI doomer cult is just another arm of the same "rationalist" philosophy that the oligarchs are using as cover for their monopolization of compute resources.
Anthropic's LLMs are a derivative work of my copyrighted material. As a joint author I hereby authorize everyone one earth to distill from them to their hearts content.
It's hard to know for certain, but that doesn't mean we should throw up our hands and say anything goes. We can still say that some things are more likely to be useful than others.
The idea that FFT's decomposition is non-optimal is surprising. What the new bound means is we basically were looking at the problem and its structure incorrectly, and therefore implicates other problems for which we have used recursion to get to our currently-best bounds.
Almost no interesting math problems are just sitting there alone in their little corner. Getting a hint like that is a big deal.
Is it surprising though? If you assume constant time when it doesn't exist (in this case arbitrary precision arithmetic), it should be obvious that you will be able to devise algorithms that are faster than the commonly accepted best results.
If you can do arithmetic on arbitrarily large integers in constant time, you can sort lists in linear time.
Because the MD length extension attacks only show up in a fairly limited set of cases and have mitigations which are considered bog standard (e.g. HMAC). If you're doing something where a MD length extension attack would harm you ... you're already lost in the dark and are going to be eaten by a grue-- avoiding SHA-256 will not save you.
Swapping out sha256 with something else when you would have had an extension attack means you're probably going to flub domain separation or miss committing to segment lengths or something else that is an inherent footgun that can't be misuse resistance your way out of but which harder to get right and more easily consequential than avoiding MD extensions.
Misuse resistance is important for sure, but the MD length extension is not in the same fragility class as stuff like 'relied on the caller to provide randomness for ECDSA'. And the hardware implementations of SHA256 have quite good energy efficiency and the algorithm overall has held up exceptionally well.
This is dead on arrival just like their, very good, powerpoint building product.
Nobody in the western world wants their value added closed source services (submit your H2 strategy deck to their cloud? Your contracts? I think not) especially when they are priced increasingly close to frontier.
Z.ai doesn't fully appreciate the level of distrust by their clients from western markets for simply being Chinese origin.
They do have credibility but it comes from being both open and having extremely capable models.
Removing one of these pillars makes a marginal risk / reward tradeoff probably a nonstarter.
Maybe LATAM, BRICS, Middle East are more willing to accept a closed / relatively expensive value prop. But not US / EU.
I mean forget the international intrigue angle. Some jackass troll with some money to burn can file a frivolous lawsuit against you seeking discovery of this material. Either the cost of production or the risk that they can find something embarrassing to pull out of context (or maybe even something debatable as wrong doing) means you'll pay some extortionary settlement to make them go away.
Beyond the obvious possible profit motive these attacks can be arms length arranged by competitors or people with an axe to grind to trip you up and slow you down. Because your activity is valuable to you there is an asymmetry-- they can bring the action cheaply because they aren't risking much, but your defense is expensive because if you flub it you could lose everything.
Part of the value of a legal advisor generally and particularly something like a legal AI is that it lets you map risk surface: "What if we started selling all the customer secrets? would we go to ultra-prison?"-- you don't ask because you're going to do it, you ask to perform a sensitivity analysis, to know where the dragons are hiding, and to setup contingencies. For stuff you're actually doing/going to do you want advice backed by a team who can defend it if it goes wrong.
All of hosted AI is a big risk from the perspective of discovery-- but particularly or legal advice/research that seems extremely radioactive and a bad tradeoff.
If anything being outside of the west might be a slight asset on this risk-- as there is some chance that they respond to a subpoena with "bite me"... but that probably doesn't overcome the other concerns you raise.
The open weight Thompson Reuters finetuned qwen hallucinates like CRAZY. ... if corporations and governments start using it it's going to cause some serious harm.
I don't agree. Asimov's robot stories are mostly about the unexpected (and usually) about the unintended and usually negative consequences of these mechanical absolute rules. They aren't being worked around, they're being followed.
Presumably they don't provide the same discount to cash because the real return on the app is stealing your data, tracking your location, etc. -- perhaps even reading out your other bank transactions and balance given that using ACH often requires giving some service the login credentials to your bank.
Have you ever bought gas? The price advertised is the cash price, if you pay by card it’s $0.10 higher per gallon. They’re giving you the cash price in this scheme.
reply