Yeah, it’s easy to think that making a biological weapon is as easy as just following a series of steps, but it’s not. It’s going to require labs and source materials that are going to be hard to get without somebody noticing. State-sponsored groups already have access to these, and you better believe they already have at least looked at weapon-building scenarios or small-scale prototypes. Isolated actors are going to have a much harder time, even if they have PhDs.

Sort of. It’s clear that Tristan and Levant were working on NS and had generated a preliminary proof that was still being verified. They were taking their time to cleanup the proofs so that they could be presented elegantly, but then OpenAI gained access to their unpublished work, evidenced by the wording of the prompt they claimed to have used, and rushed to scoop them. This kind of shit happens all the time in academia. It’s rather unsurprising that it would also happen between OpenAI and Anthropic.

Which in itself is appropriate (most employees of companies have a responsibility to protect trade secrets from competitors), but the problem here is they wanted to take credit for getting to a solution faster than their competitor, which they did using publicly non-disclosed information.

And because OpenAI is desperate to show that their models can generate proofs to hard problems, this scenario is highly likely. They were augmenting their training data with examples of partial proofs and “Oh look, there is a nice draft of a paper that someone was editing with Codex”

They also heard that someone at Anthropic was working on this, so they decided that it was imperative they publish first.

If you read Tristan’s actual statement, which someone posted a link to below, the reason he is suspicious is because the details are of the approach, including the specific wording used in the prompt, were non-obvious. That and the timing of when the work supposedly started circumstantially points to theft of Tristan and Levant’s work. Tristan does not actually accuse of OpenAI of theft because he doesn’t have any proof, but the sequence of events is definitely suspicious.

There is also the issue of the pressure tactics: respond to a meeting invite the same day, respond to a meeting invite on Sunday of Labor Day weekend, release your proof so that we can follow up with ours the next day. They clearly were rushing their result because they wanted to scoop anything that Anthropic may have been working on. Meanwhile, Tristan is more interested in the academic discourse, but he can no longer release the result he wanted to release, because writing good proofs takes time and OpenAI didn’t want to give him that time.

Sorry, uh what? Bohr’s model is not useful for anything beyond hydrogen. That’s why high school chemistry classes only teach it as the first step in the historical development of atomic orbital theory.

I don’t disagree, but I will go back to the point I was trying to make previously. Is the limitation one due to perception or decision-making? If the problem is you can’t see enough of the right things, that supports the argument that you need more sensors. But if the problem is that you aren’t making the right decisions with the information you have, more information is not necessarily better.

If performance is limited by object detection, then yes, but if performance is limited by decision-making and contextual awareness, then not necessarily.

That is not what I said. I said Waymo’s approach to self-driving is a sunk cost. The difference between the Waymo and Tesla approaches is not just sensor count. Waymo chose their approach a long time ago and are locked in to it. Tesla is also locked in to their approach, but they started later and framed the problem differently. It’s more complex than just lidar vs no lidar.

Perhaps, but again, if perception (ie ability to see) is the primary factor limiting safety performance, there is more than one way to overcome that.

Objectively, the Waymo driver operates in a much smaller service area than most humans drive, with the majority of miles accumulated on a small set of curated routes. If “safety is solved”, as you claim, why doesn’t Waymo operate everywhere that humans do?

You are bringing up engineering trade-offs. LiDAR brings some advantages, yes, but it also comes with its own disadvantages. The 5th generation I-PACE uses 34 independent sensors. So the real engineering trade-offs are much more complex than your example of a camera vs a thermocouple to measure temperature.

How about we get as good as humans first, and then we can talk about being better than humans. Sorry, even Waymo is not there yet. The claim is that safety performance is due to a lack of being able to see and more sensors is the solution. If we take that statement at face value (ie lacking data), then additional cameras could just as easily overcome the limitation as additional sensing modalities. But most accidents caused by humans are due to judgement errors, not an inability to see, so one could just as easily argue that super-human performance should be achievable without additional sensors. Incidentally, all approaches to self-driving (including Tesla’s) use way more sensors than humans possess, so if more sensors is all that is needed, we should have achieved super-human performance a long time ago.

Look, let’s be straight. When Waymo started their experiment in self-driving more than 20 years ago, cameras were not as good as they are today, embedded computers were not as powerful, there was a limit to how much custom hardware design they could do, power consumption was an issue, and computer vision was in its infancy. The only feasible approach at that time was to use a rich sensor array coupled with large data heuristics—maps that tell the car which lane to be in when going through an intersection, which light corresponds to the lane, which intersections have stop signs or yield signs, which intersections have no turn on red, what is the speed limit, which intersections should be avoided entirely because they are too complex, etc—to lower the burden of perception and decision-making that had to be made directly by the car in real time. In other words, the current Waymo driver and all of the engineering that was put into it is a sunk cost that they are not going to back out of now. If they were starting fresh today, I would be willing to bet that they would not take the same approach.

The “dashboard” isn’t publicly available so there is only the posted screenshot to go off of, but it literally looks like it converted a pdf into a set of slides and that’s it. The fact that these posts appear to be nothing more than an advertisement for Microsoft products, coupled with Satya not knowing what skill he was actually using (/grill-me not /drill-me) makes me think he asked the chatbot to design a dashboard, not build one. So it happily generated a bunch of buzz words (and a mock screenshot courtesy of the pdf) on slides and name dropped a bunch of Microsoft products just for the shareholder meeting. In other words the only actual work done here was the report generated by the, presumably human, Morgan Stanley analyst.

As someone else said, “The illusion of value.” On the other hand, consultants get paid millions to generate equally vacuous outputs, so maybe there is some value here, to the people who consume PowerPoint slides.

This is a meaningless statement. I too can create a tool to generate a list of peptide candidates with minimal effort. It may even be somewhat useful if it based peptide sequences on homology searches or some other relevant biology instead of random string generation. This has been an active area of research for than 20 years. In order for this to be newsworthy. Claude has to be better than what already exists. How many novel candidates does it generate that actually have useful antibiotic properties? Do I have to screen through a list of 100 candidates to find one that actually works? If so, that’s not much better than a BLAST search and it costs a lot more. What is the strain selectivity of the new antibiotic? Is it broad or narrow spectrum? How easy is it to manufacture? Are there any toxic side effects?

Assuming a new antibiotic is actually what’s needed, instead of using one of the many beta-lactams or combination therapies that already exist, generating the candidate is the first and easiest step of a long and expensive process to developing a novel drug.

It should be fairly easy to carve out provisions for exceptional circumstances (ex: bankruptcy proceedings), and is probably common practice already. But I agree the language has an implied target which is not good for getting legislation passed, generally. It should be possible to propose reasonable accommodation for majority circumstances without sounding like a manifesto.

That will run into problems with federal copyright laws and other IP protections. While it sounds just, it is unlikely to get anywhere. More likely is large companies would have to have an end-of-life plan for their product (which they likely do already) that includes considerations for how to keep the game playable after servers are taken offline. It could be as simple as releasing the protocol to enable third parties to host servers. Or, better, designing the game for offline play from the beginning so that the final patch set is just a simple switch.

Regardless of the outcome, however, this really does seem like a pre-eminent example of “first world problems”.

Well, it says in the abstract (and they have further detailed calculations in the paper) that they achieved an average desalination rate of 1.76 kg/m^2/h. So that’s about 1 gallon of water in one hour with 2 m^2 panel. That said, they did their tests using a 9 cm^2 panel, so yields may change significantly on scale up. Not sure if that translates into meaningful cost savings at scale, but it does seem like a significant advance in solar desalination technology. It seems like the paper was focused mostly on salt harvesting, though, rather than desalination. So I expect the real utility to be the combination of the two rather than desalination by itself, if it scales.