If he was French, this would be an attaboy moment - without the spy allegations, he'd be cheered on. However... in such a puritanic country as US, I suspect it was the second disclosed affair that cost him his position.
After all, it changed the narrative from Huang being "someone who fell to nation-state honeytrapping" to a regular, equal-opportunity filanderer.
Using data feed to guide pricing and set initial figures: legal.
Agreeing, either implicitly or explicitly, to let the pricing data feed set the price floor: collusion.
Leaving units unsold and refusing to lower the price despite insufficient demand: usually a bad business decision, possibly indicator of a market failure.
Leaving units unsold to maintain agreed-upon price floor despite insufficient demand: collusion, market manipulation. Congratulations, you are a cartel.
You're assuming they were the first kind of fools rather than the second kind.
> Leaving units unsold to maintain agreed-upon price floor despite insufficient demand: collusion, market manipulation. Congratulations, you are a cartel.
Which is the thing that doesn't work when many other landlords aren't in on it, because then they get to fill their vacancies while you're stuck with the empty units.
Notice also that antitrust prohibits attempts to monopolize. Fools are quite capable of doing something which is still a violation of antitrust laws despite being ineffective and unprofitable.
I'm not sure this is specific to Claude, because I'd expect every LLM to do the same here. This behaviour likely goes under "what did you expect to happen?" - if the service describes its intended semantics and the agent follows that description, is that misalignment or something different? (To turn the thing around: would you as a user follow the documentation without any further context?)
I get the feeling that alignment problem might well be unsolvable and this thread is an excellent example. The world is not going to run short of contradictory problems due to gazillion edge cases... and to solve for them we are apparently now building a new layer of expert systems[0] to somehow tell the agents what they should do in every conceivable circumstance.
Only in this case the decision trees all have one common requirement: "make no mistakes"
> This behaviour likely goes under "what did you expect to happen?" - if the service describes its intended semantics and the agent follows that description, is that misalignment or something different?
Something different.
Don't give the AI access to data you don't want it giving out. Any AI you use is no better than some fresh out of college intern at best. They see 'email address required' and they put the email address into it because you instructed them to do a task that required passing over your email address and if you didn't want them to do it why would you have given them that task?
If there is any other issue here it would be regulatory. I could see the EU framing this as the AI company illegally sharing your PII information if it passed your email address to another company without your explicit permission.
This is not directed at you personally but Anthropic (and everyone around this debacle) in general: please put some real effort behind planned crisis communications. The reporting on this has been atrocious and the comms coming from Anthropic have not really helped.
FWIW: the decision to publish an overriding version bump across all lanes was the correct one. It simplifies comms and helps to answer the first question every affected party has: "but what does this mean to ME?" - much easier to deal with the flood when you can start by sending back "are you running, or have you ever installed version Y?"
Time for technically accurate nuance comes shortly after, and in this case it has been lacking. Yes, you are dealing with confusing mix of various extension marketplaces. Explaining that to a questioning audience who do not have the necessary background information is non-trivial. Kevin's blog is pretty well known, but in this case the real updates are only coming in via a separate Mastodon thread (itself polluted by uninformed commenters) so the real, tangible and understandable details have still been missing.
Claude and its surrounding ecosystem are a huge, high-value supply chain compromise target. Cursor's extension being compromised is not your fault but it is still your problem.
Disclosure: done enough security incident comms in high-stakes environments with extremely demanding clients, operating under various regulations.
> But Anthropic (and everyone around this debacle) in general: please put some real effort behind planned crisis communications. The reporting on this has been atrocious and the comms coming from Anthropic have not really helped.
Anthropic is incapable of communicating. They rely on random employees in random threads on random social media to do some kind of PR and posts. Regardless of the issue. Or even non-issues but features.
You're confused. It's not Anthropic's responsibility to monitor and fix everything that's said correctly or incorrectly about them. They're doing excellent given what they're capable to do in this situation.
You run a business and have millions of high paying customers. Someone happens to use your brand to market things and create harm for the society. Legally, you may be clean. But, your first line of defense is not "it is not my problem".
Yeah, but every now and then someone working for Anthropic tweets something, everyone quotes it as gospel, but there's no official statement anywhere on their website. They have the worst communication ever, it's one of the things I like the least about Anthropic.
It’s strictly speaking not their responsibility, but it is something that they should do if they want to help their customers.
In general, they need to make a really big post somewhere that’s not a HN comment that this was in fact a compromise of the Cursor marketplace, not an employee’s machine, because they’re getting a lot of (unjustified) bad PR.
No but their no doubt highly paid lawyers should have sought relief from their trademark being infringed, and demanded the extension was pulled long ago.
They didn't because they were benefitting from the situation. (Users got to use their preferred tooling.) Until they weren't. That makes it more their problem, in my mind.
Sounds like innovator's dilemma to me. Their current production lines might be too profitable[0], so diverting expansion capacity to a yet unproven technology (to Micron's to-date profit margin requirements) might be too big of a financial hit. Since they're the ones who came up with 4F design, they must hold a ton of patents on the various incremental improvements. At which point, from a purely cynical and utilitarian view, it would make financial sense to let other players polish the remaining 20% and use patent portfolio as a wedge for mid-term technology transfers.
Or in short: "let others figure out the technical complexities and surrounding economics." - Which would make sense from a company that believes they can catch up once the expensive leg of the manufacturing/scaling research has been completed. Whether that bet is good or not, we'll see.
0: Didn't we read a few days ago here that all of 2027 production capacity has been sold out already?
Given that the stated plan of Micron is to focus on 3D scaling, keeping per-mm^2 production costs low makes perfect sense. When Samsung introduced their first volume V-NAND flash, the 2D feature sizes were larger than leading edge NAND flash because of the tolerances inherent in producing 90 layer flash (aspect ratio matters!) as well as the cost of production equipment. That trend has continued as it is cheaper to produce flash with more layers than to compensate for difficulties arising from decreasing amounts of charge kept in ever smaller cells.
In layperson DRAM-speak this translates to the following: rowhammer attacks are mitigated at the hardware level by increasing the size of and spacing between adjacent cells. At some point further feature size scaling makes it no longer be possible to produce 4F cells that are sufficiently robust against rowhammer type attacks to meet Micron's internal reliability targets. This would be akin to when Intel hit the power wall with the Pentium 4 in the early '00s forcing a change in engineering direction away from merely scaling clock speed.
With Micron's insight into the DRAM order pipeline, management likely has to factor in the chance of oversupply in the next couple of years as significant, which makes the shift in direction even more sensible. DRAM price crashes are a regular occurrence in the industry which lead to anti-dumping tariffs and legal battle in the 1990s (as well as Intel's shift away from DRAM towards complex logic devices in 1985). I'm pretty sure Micron's management is consciously working to avoid a repeat of that protracted legal battle.
I'd rather have an engineering problem over a legal problem any day of the week.
> Which would make sense from a company that believes they can catch up once the expensive leg of the manufacturing/scaling research has been completed. Whether that bet is good or not, we'll see.
The article mentions CXMT has committed to 4F^2 server RDIMMs by end of 2026, and a linked article mentions that CXMT has halved the fab build time from the industry-standard 2 years to just 1. How would Micron catch up with that?
The memory triopoly is in the midst of a slow motion wreck because pumping the brakes will reduce profits; the incumbents are going to have their butts handed to them by CXMT in 2-5 years, and they'll all act like they didn't see it coming as they are bailed-out by the US or SK government, for "National Security."
There was a rather interesting article (sadly LLM-written and not very well either, because the underlying content would be actually good) about this era being "AI's 2 FPS moment".
The comparison was made with the graphics engine team at Microsoft who were writing the first version of DirectX at the time when then-current hardware could just about do 2 frames per second. The sentiment apparently was along the line "why would you bother?" By the time the release grade stack shipped, hardware had caught up and 30 FPS was commonly doable.
We should expect to see similar step-change improvements over time, as the algorithms are refined and hardware further adapted to the new bleeding-edge constraints. In a way we already are, because apart from the frontier models ballooning like crazy, the "regular" class models are becoming more capable in smaller sizes.
Maltese thick rabbit stew (think: ragu) is delicious. And as mentioned in a parallel thread, hare is fairly common in Finland as well.
The meat is gamey and indeed very lean, but when you slow cook it to the point where it falls off the bone... yum. Tends to work better with roasted potatoes than mash.
This has been going on for FAR longer than LLMs have existed. We call it "internet background radiation", because it's everywhere; it usually shows up within minutes of a host coming online. Or just about immediately after your first LetsEncrypt certificate gets issued.
Pretty much every infosec company of non-trivial size runs one set. The old venerable services are Shodan and Censys. Then there are a various security "researchers" who like to point their crapware vulnerability "scanner" at a subnet or ASN and let it rip.
And then there are of course all the actual botnets and low rent access broker wannabes trying to pop shells where ever they can, so they can resell that access.
Also known as the trading model's shelf-life. Based on what I have heard and learned, the average active trading model has a useful life of around 18 months. After that the rest of the market has adjusted and its edge is gone.
If rumours are to be believed, several "slow" hedge funds have models that remain useful and profitable for 5 years or more. Then again, those models are not used to conduct exchange trades but rely more on aspects of fundamental analysis.
Rather curiously some of the largest banks and tech-heavy asset managers tend to sponsor meetups and events like PyData - every now and then doing a slot on some of their open-sourced stuff. You might see a talk on 5-year-old trading model internals with large chunks of it opened up, or they could present some of the internal UI or visualisation libraries. The latter tend to be things that they no longer actively develop (they're "ready") but maintain for their ongoing persistent needs.
Of course anything that actually brings them money and/or gives an edge is not even discussed.
Honestly: People think that "trading models" (whatever this should be) needs to be somewhat "supersophisticated" and "extreme driven by whatever complex math" - the brutal truth, esp. for smaller trading shops is much more simpler:
- Standard approaches like trend following or mean reversion are working very well since decades. (I can speak only for the last ~25 years)
Complexity in trading is not about "building that one specific niche-super-strategy", but more about putting all the ideas in a reproduceable approach/process to repeat it over and over again to grow the money.
> You can agree with the rules anthropic wanted, but having rules set by a private company at all that apply to the military does seem fair for the military to object to.
Oh the hypocrisy and irony. This coming from a country whose military sets rules on other countries' armed forces how weapons systems bought from US can be used.[ß] By this standard US itself should be designated as a supply chain risk for everyone else.
ß: for a very long time in the Ukraine "conflict", US refused to allow their gear from being used to attack targets outside of Ukraine's borders.
If he was French, this would be an attaboy moment - without the spy allegations, he'd be cheered on. However... in such a puritanic country as US, I suspect it was the second disclosed affair that cost him his position.
After all, it changed the narrative from Huang being "someone who fell to nation-state honeytrapping" to a regular, equal-opportunity filanderer.
reply