troyastorino

Tokens do have a clearly calculable intrinsic cost. There's the marginal cost of production (i.e. the inference cost) and the amortized R&D cost that goes into the model producing them.

Yes, value is hard to calculate, but luckily market pricing mechanisms exist exactly for this purpose. There isn't a better number to use than what people are willing to pay for them.

So he's saying that on an enterprise plan, he'd be spending $2,180.16. He's not paying that much, but enterprises are.

Matt Levine often says something like "insider trading is not about fairness, it's about theft." The problem isn't that it's less fair to some stock traders than others, and that stock trading should be some form of perfect gambling where everyone has an equal chance of success. Stock trading is inherently about exploiting information asymmetries — that is what all "non-insiders" are trying to do. But insider trading is wrong because it's effectively stealing confidential information from the company & shareholders, which is in violation of & conflict with the fiduciary responsibility that board members, executives, and employees generally have towards shareholders.

Conceptually, I think that is the right analogy to think about. Prediction markets "want" to be a more accurate source of information, just like stock markets, so from that lens "getting" information to be more accurate is good. When government officials are placing bets on prediction markets, though, it's a massive violation of operational security, and leaking confidential information. They probably think that they are acting anonymously, but it creates so many opportunities for unfriendly state actors to get information, especially if people do it consistently.

The point of a humaniod is compatibility with systems that have been built for humans, which is...nearly all systems that have been built. Environments and tools that exist in the world are ones that have been built for humans, so for a single robot to be able to interact with almost all of those, that robot needs to be shaped like a human.

The overlap between the Effective Altruism community and the Rationalist community is extremely high. They’re largely the same people. Effective Altruism gained a lot of early attention on LessWrong, and the pessimistic focus on AI existential risk largely stems from an EA desire to avoid “temporal-discounting” bias. The reasoning is something like: if you accept that future people count just as much as current people, and that the number of future people vastly outweighs everyone alive today (or who has ever lived), then even small probabilities of catastrophic events wiping out humanity yield enormous negative expected value. Therefore, nothing can produce greater positive expected value than preventing existential risks—so working to reduce these risks becomes the highest priority.

People in these communities are generally quite smart, and it’s seductive to reason in a purely logical, deductive way. There is real value in thinking rigorously and in making sure you’re not beholden to commonly held beliefs. But, like you said, reality is complex, and it’s really hard to pick initial premises that capture everything relevant. The insane conclusions they get to could be avoided by re-checking & revising premises, especially when the argument is going in a direction that clashes with history, real-world experience, or basic common sense.

The answer is a little nuanced.

We train on real records, and even though they are de-identified in training we still have to keep the model closed and under careful management to protect against the possibility of information leaking.

We are, though, definitely invested in this corner of research, and want to be able to work with others to push medical AI forward.

Given that, the best model for us is to collaborate on an engagement-by-engagement basis. For now we'd look to find ways to do the work directly involving LLMD within our systems.

If you research in the field and have some ideas, I'd love to chat!

(Co-founder of PicnicHealth here; we trained LLMD)

Accuracy and deploying in appropriate use cases is key for real world use. Building guardrails, validation, continuous auditing, etc is a larger amount of work than model training.

We don't deploy in EHRs or sell to physicians or health systems. That is a very challenging environment, and I agree that it would be very difficult to appropriately deploy LLMs that way today. I know Epic is working on it, and they say it's live in some places, but I don't know if that's true.

Our main production use case for LLMD at PicnicHealth is to improve and replace human clinical abstraction internally. We've done extensive testing (only alluded to in the paper) comparing and calibrating LLMD performance vs trained human annotator performance, and for many structuring tasks LLMD outperforms human annotators. For our production abstraction tasks where LLMD does not outperform humans (or where regulations require human review), we use LLMD to improve the workflow of our human annotators. It is much easier to make sure that clinical abstractors, who are our employees doing well-defined tasks, understand the limitations in LLM performance than it would be to ensure that users in a hospital setting would.

Gas taxes are a direct mechanism to price the negative externalities. Yes, gas taxes also cover road infrastructure, but the amount of gas purchased (and therefore taxed) is directly proportional to the negative externality of CO2 emissions.

Google groups is also the mechanism for defining groups for viewing/editing permissions, comments, calendar invites, etc across Google Workspace and for access permissions in GCP. That group definition aligns nicely with being a mailing list grouping. Those use cases are not going to go away, and I would expect Google will invest in them more (as Workspace has been an area that Google has significantly invested in, and I'd argue has been doing a good job).

I’ve had mixed success typing things verbatim vs hand writing verbatim.

When learning to program, I found typing out code samples had a similar effect to writing them out by hand, and felt more natural to the domain. But in other subjects in school, it felt like my comprehension was lower when I typed lecture notes instead of handwriting them. I was never sure whether that was something inherent to hand writing vs typing or it was just because I had more of a barrier between typing and thinking than I did with handwriting.

It always seemed clear that the value of writing something verbatim was that it forced your brain to internalize it in a way you don’t need to when reading. It takes longer to hand write than to type — maybe that is the value of hand writing vs typing? Or when typing I needed to think more about the act of typing, whether that was thinking about formatting/positioning or just being less natural than handwriting? Maybe programming is just a special case where you always think through typing?

I lean towards thinking it’s somewhat domain specific. Clearly with physics or math if you had to typeset equations with LaTeX that would get in the way, and typing would be a barrier to understanding. With programming, every idea is expressed through typing, and so typing is a natural way to imprint ideas on your brain.