HN user
I agree you probably don't need top-dollar bay-area engineers for this, but hardcore outsourcing to a LCOL probably isnt going to work either due to novelty and generally being setup to do the more rote thing (generalizing a ton here obvi). This feels like something in the middle.
This, taken in combination with the SAE paper, the golden-gate claude paper, the feelings / introspection paper, and note in the fable system card (that they are silently nerfing responses about activation shaping), is basically confirmation to me that they have a new technique they they are using during training (along the vibe space of these mechinterp papers), and its probably some kind of representation learning akin to the core ideas of JEPA.
(Nb: not an expert / in the labs, just opining)
If you know whats going on, could you link to em?
Somewhat similar to you, I also got the sense that this was a pointed response/positioning to some controversy. Not because I have any clue whats going on (I dont), but just cuz it pattern-matched with the style/tone of a PR statement.
Frontier labs have their own variants of MLA and certainly their own balance/scaling-laws for things like MoE vs FC vs Attn. MoE scales really well for inference with horizontal scaling + batching, which these guys luv.
On the architectures side, I'm a lot more interesting in attention residuals than anything else, one of those things that seems obvious in hindsight and Kimi have proven it at scale.
Yes. A tax on the ultra-wealthy, rather than a measure aimed at increasing housing.
Its very roundabout as NYC can only make taxes for NYC, but the net aim is to increase the effective tax rate for the ultra-wealthy, using secondary property as a proxy for that.
Edit: AND WE (I) LIKE THIS because progressive taxation is the core play of fixing income/wealth inequality
Theres no gotchas in the quoted text, but curious if you got the impression from your emails that any of it was.. AI generated?
The camber, affirmation, word choice, triplet phrase... leaves me wondering. But without a smoking gun its hard to know if a model call was fired.
In this case you can say "the contract we reviewed was poisoned via technical means to show different words, depending on how the file was read". Perhaps if pushed you can say you loaded it into GCAI/Harvey/Legora and read/reviewed it there.
There is no parallel construction where this wasn't deliberate & malicious, so it seems really high risk given the judge would rip you a new one if discovered.
You can't rely on the defense that the other party didnt read it, if you made it show different words depending on how it was loaded.
It seems like the main attack scenario for this + legal AI would be during discovery: if opposing counsel gave you a poisoned PDF, and you threw it into one of these products to help you sift through it and got bad answers.
However, wouldnt this be a rather risky move? Courts authorized the discovery, so I imagine the judge might loose their marbles and throw the hammer at them if this came to light.
This is a great writeup! thank you!!
Reminds when i did noogler training back in the day and one of the talks described a cascading failure at a datacenter, starting with a cat which was too curious near a power conditioner, and briefly conducted
They absolutely have backups, I presume they were ineffective or also down for _reasons_.
Worth adding also that you can only use these vectors to corrupt the page cache for files reachable in your mount namespace.
Usually with containers, almost nothing is shared with the host namespaces (tho likely shared with other container namespaces, hopefully none of those are --priv).
Nah, even insane token costs don't come close to the costs of labor.
Most likely this is just 'AI-washing' - dressing a layoff for economic reasons (such as propping up their shrinking margins) as something more palatable to investors (AI).
Between this, the emotions paper, golden gate claude etc, it doesn't seem like such a stretch that Anthropic are doing some kind of activation steering as part of training (and its part of their lead)
Theres been multiple papers discussing how only updating parameters that have high agreement in update direction leads to less overfitting and better generalization. Lemme see if I can find em.
https://arxiv.org/abs/2411.16085 - set updates to 0 where theres disagreement in the sign of the parameter update - got accepted!
https://arxiv.org/pdf/2412.18052 - discard gradient updates from batches/minibatches that disagree where disagree means cosine distance threshold (they solved for 0.97 or something being optimal)
This attack class lets you escalate from any user to UID 0. Not running as root won't save you, in fact, this attack is for those processes not running as root.
However, if you are in a user namespace where UID 0 doesn't map to system-wide capabilities, and you dont share page cache for the setuid binaries on the system, this attack doesn't lead to LPE.
That is, if the batch signal on a parameter exceeds its leave-one-out noise, update it; if not, skip it. This is a one-line change to Adam that accelerates grokking by 5x, suppresses memorization in PINNs, and improves DPO fine-tuning, eliminating the need for validation sets entirely.
Does anyone understand the formula they expressed above this sentence? is this just the classic "skip updating parameters with high gradient/loss variance in multiple batches/samples" ?
my guess is they want to do AI/O as part of their event loop explicitly, and blocking a thread in a syscall waiting for an IOP (ala std::fs) isn't the vibe.
high surface energy?
mahm/father I yearn for the spicy surface
Unless "with experience in the field" == academia, disagree. In particular I remember the early discourse & hype around Wireguard, it was discussed as if perfection was an achieved outcome.
The archive-handling code was in lean-zip, it just seems the verifiers forgot to write proofs for it (still a bug).
Thats not the main finding of the article however. The main bug found was actually in the lean runtime, affecting all proofs using scalar arrays where the size of the array is not bounded.
Is it fair when it comes to formally-verified software to only consider bugs that violate a proof, and ignore everything else?
Formally-verified software is usually advertised "look ma no bugs!" Not "look ma no bugs*" *As long as this complicated chain of lemmas appropriately represents the correctness of everything we care about.
In boating theres often debate of right of way rules in certain situations, and some people are quick to point out that giant tanker ships should be giving way to tiny sailboats and get all worked up about it*. The best answer I've heard: they're dead right! that is to say as right as they are dead (if they didnt yield) lol. In the same vein, I think someone who assumed that a formally-verified software was perfect and got hacked or whatever is going to be a bit wiggly about the whole thing.
* = Technically the rules prioritize the tankers if they are "restricted in ability to maneuver" but everyone loves to argue about that.
Looks like a normal distribution about the chaos mean to me. I appreciate its not everyones cup of tea, but I like this style of writing.
This 'fingerprint' function is super interesting, I imagine this is a signal they use to detect non-claude-code use of claude-code tokens: src/utils/fingerprint.ts#L40-L63
Strong agree on it all being a legal minefield / new grass.
But was almost certainly part of its training data, which complicates things
On this point specifically, my read of the Anthropic lawsuit was one of the precedents was that if it trains on something but does not regurgitate it, its fair use? Might help the argument that it was clean-room but ¯\_(ツ)_/¯
Is your (1) description of clean room implementation, and (2) description of what was done, actually correct?
(1): my understanding was that a party _with access to copyrighted material_ made the functional spec, which was communicated to a party without access [1]. Under my understanding, theres no requirement for the authors of the functional spec to be 'clean'.
(2) Afaict, they limited the AI to access of just the functional spec and audited that it did not see the original source.
Edit: Not sure if sharing the 'test suite' matters, probably something for the courts in the unlikely event this ever gets there.
[1] Following the definition of clean room re implementation as it relates to US precedent, ie that described in the wikipedia page.
I feel exactly this.
One thing I will add: while AI is getting really good at _doing_ the software building bits, I haven't yet seen it well integrated into the decision-making and political structure of organizations. Right now, it seems best in the hands of a high-agency individual empowered and able to make changes or 'ship' something, with them acting as the bridge.
This of course, is not a technical challenge, but I would expect the change in structure of organizations to make this more efficient to be slower than the pace of improvement we've seen over the last few years.
On (1): Might be useful to separate investment flows from the rest of US's economic activity.
AI investment is propping up capital flows, the GDP statistic, and responsible for most of the gains on SPX, but its still a small fraction of the economy.
Does Waymo have the same object permanence and trajectory prediction (combined) to that of a human?
On this note specifically ive actually been impressed, ie when driving down Oak st in SF (fast road, tightly parked cars) I've often observed it slow if someone on a scooter on the sidewalk turns to look toward oncoming traffic (as if to start riding), or to slow passing parked box trucks (which block vision of potential pedestrians)
There are settings to make it reproducible but they incur a non-negligible drop in performance.
Unsurprising given they amount to explicit synchronization to make the order of operations deterministic.
Its not deterministic. Any individual floating point mul/add is deterministic, but in a GPU these are all happening in parallel and the accumulation is in the order they happen to complete.
When you add A then B then C, you get a different answer than C then A then B, because floating point, approximation error, subnormals etc.