I also write at https://splittinginfinity.substack.com/
harsimony
Karma: 605
Thank you for offering a more constructive comment.
We did a power analysis to set the total number of trials (iirc assumed d=0.5, alpha=0.05, 80% power, so ~30 total test weeks and 10 weeks/person). However, the design proved unsustainable for us and the Fitbit dropped one persons data.
Though in some sense it worked out, we can pursue a better trial now.
Thanks for this background, it’s super helpful!
For dosing, I think we based the dose on the Deadwyler et. al. study in monkeys as well as user experiences.
Unfortunate that a 10-20x dose in humans seems to have small effects? I would have expected them to use a C-terminal -NH2 modification on their Orexin-A to prevent degradation, but it doesn’t look like it (but maybe they did and I missed it). If that’s the case, might reduce the effective dose gap between ours and theirs somewhat.
I’m excited to see our results at a higher dose, though part of me is frustrated by how difficult peptides are to work with. But hopefully Takeda or someone else will perfect small-molecule orexin agonists!
Yeah we used C-terminal NH2 modified orexin to prevent degradation but its possible it simply wasn’t effective.
Interesting that orexin sprays haven’t been working, I’ll have to look into this. Do you know the names of any off the top of your head?
Love that post! Made me realize that sleep need reduction therapies have to be pretty specific in what receptors they hit. A stimulant like modafinil that hits orexin + other stuff doesn’t reduce long-term sleep need in healthy individuals right? So a sleep need therapy needs to stimulate for just the right window, while enabling efficient sleep at night.
Oops, that should read “Sam preferred to get up early …” my bad!
Interesting! Once you get a good nights sleep or a break does your performance go back to normal? Or does it take a few days?
Null Results From An Orexin RCT
I found a thoughtful post on this topic that runs counter to your argument: https://jasher.substack.com/p/is-murder-falling-because-of-improved
Aggravated assaults are simply too poorly reported historially to draw firm conclusions about the long term. I would guess that improvements in trauma care have likely helped make shootings more survivable today than they were 50 to 60 years ago. But a lack of standardization of reporting practices among law enforcement agencies in the 1960s and 1970s and little to no data on shootings until relatively recently lay doubt into whether more people actually survived shootings in the 1960s and 1970s than today.
My takeaway — as is frequently the case — is that people should avoid reading strong conclusions into crime data that can be explained by changes in reporting practices. Murder is falling now because fewer people are getting shot. Going farther back to evaluate the impact of improved trauama care relative to the 1960s is, in my opinion, not advised.
See also this image from Skaran et. al.:
My overall take is that:
Lethality of GSW’s is roughly flat in recent decades.
Because of 1, crude murder rate in recent decades is a reasonable proxy for things like assault, attempted murder, murder, etc.
Violent crimes have probably fallen over time.
Comparisons to the 60′s-80′s are fraught because of underreporting.
Crime data is messy, but I think when you stack various measures of disorder, violent crime etc. and try to adjust you see that crime is in fact falling. This is the best explanation for why we see flat/falling crime across different categories, in many places, under different reporting regimes, and with different levels of anti-crime tech. This also fits with data on how teens are drinking less and driving less because of the internet.
This makes it hard to validate peoples vibe that everything is awful everywhere. I think this vibe is better explained by social media over-representing crime and disorder rather than actual increases in crime. There is good evidence of increasing negative polarization in the media in recent decades:
https://www.astralcodexten.com/i/179630044/conclusion
Negative polarization can explain why we have vibescessions, crime waves that don’t match the statistics, and frequent moral panics about non-issues like data center water use.
This is interesting! I think it would be worthwhile to check the robustness of your conclusions by focusing on e.g. modular addition and trying larger/smaller models or changing the weight on the regularization term.
The spikes can be understood using centralizing flows, you may want to plot the top Hessian eigenvalues to clean the plot up:
https://centralflows.github.io/
Feedback on the plots: I would find them easier to read if training and validation were the same line. I.e. dashed lines at beginning for train, switches to solid lines for validation so you can see the full path of the model.
Some links you might find useful: https://www.lesswrong.com/posts/jusq6kyZ6XSrtW3Bf/paper-summary-omnigrok-grokking-beyond-algorithmic-data
One thing to recognize is that the cost of renting capital is the real interest rate. So for people owning capital, a higher real interest rate is good. But, as more and more people have capital to loan out, real interest rates should fall (because people compete). This isn’t a cost disease per se.
An intro to the Tensor Economics blog
I would count your consulting service as RLaaS essentially. I’ll admit, RLaaS is a buzzword that obscures a lot. “Have AI researchers and domain experts iterate on current AI models until they are performant at a particular task” would be more accurate. Things I think this model will apply to:
Anything involving robots. Consider the journey to self driving cars with lots of human data collection, updating the hardware, cleaning the dataset, and tweaking algorithms. Any physical manipulation task that has to be economically competitive will need a lot of input from experts. Factory managers will need robots that operate under idiosyncratic requirements. It’ll take time to iron out the kinks.
To a lesser extent, repetitive internal company processes will need some fine tuning. Filling out forms specific to a company, filing reports in the local format, etc. Current LLM’s can probably do this with 90% success, but pushing that to 99% is valuable and will take a little work.
Research-heavy domains. The stuff covered in publications is 10% of the knowledge you need to do science. I expect LLM research assistants to need adjustment for things like “write code using all these niche software packages”, “this is the important information we need from this paper”, “results from this lab are BS so ignore them”.
My priors are that reality is detailed and getting a general purpose technology like modern AI to actually work in a particular domain takes some iteration. That’s my key takeaway from that METR study:
https://www.lesswrong.com/posts/m2QeMwD7mGKH6vDe2/?commentId=T5MNnpneEZho2CuZS
These are good points. I’m uncertain about what models will form the foundation of RLaaS. But I think your point about where the task-specific data teams are working is more important. Off the top of my head, I think there’s 3 bins:
For a lot of programming tasks, big AI companies already have lots of expertise and users in-house, so I expect them to dominate production of code generation.
For some tasks like writing marketing copy, LLM’s are already good enough at this. There’s no business training models further here.
Most interesting are tasks that require lots of tacit knowledge or iteration. For example, getting to self-driving cars required a decade plus of iterating on algorithms and data. I imagine lots of corporations will privately put a bunch of effort into making AI work on their specific problems. Physical tasks in specialized trades are another example.
For tasks in #3, the question is whether to join up with the big AI companies, or develop your own solution to the problem and keep it private.
RL-as-a-Service will outcompete AGI companies (and that’s good)
You may be interested in this series, especially the post on “three prong bundle theory”: https://www.greaterwrong.com/s/EA2uNqKjmu2NzFhRx
One good framing is to consider the rights digital minds need in order to participate in a market economy. They need property rights, freedom of speech, freedom of association, and so on. By being able to participate in market exchange, digital minds may prefer to be part of society rather than fight against it. Comparative advantage is a particularly good reason to cooperate with others.
Market rights: https://splittinginfinity.substack.com/p/markets-dont-work-without-individual
My comment on personhood and the value of being punish-able: https://www.lesswrong.com/posts/4m2MTPass3Ri2zZ43/legal-personhood-three-prong-bundle-theory#bpavKPBwJbCJtv8QA
The Data Scaling Hypothesis
Good points on an important topic. Thank you for this series.
One thing I’d like to point out is that receiving this form of personhood is highly valuable. In other words, being punishable makes you more trustworthy and safer to engage with. So AI’s and digital minds might voluntarily construct methods by which they can be punished for wrongdoing.
This right-to-be-sued is an important legal right. I covered some discussion on twitter about this in point 3 here: https://splittinginfinity.substack.com/p/links-15
The key quote from Issac King here:
If the laws are enforced on everyone equally, then people can safely interact with each other, knowing that they have recourse if they are wronged. But when one particular group of people is exempt from the law, the safest thing for everyone else to do is to avoid having any contact with that group, because they are now uniquely threatening.
The people being stolen from are not the only victims of the decriminalization of theft. The victims that nobody sees are all of the unlucky but perfectly trustworthy people who are now pariahs because society has decided to remove their ability to enter into binding agreements. To remove the social safety net that allows everyone else to feel safe around them.
Nicely written proposal, thank you.
In truth, I’m quite concerned with such a proposal being implemented. Part of this is, as you mentioned, the risk of lock in. In particular, a global entity with a monopoly on violence is alarmingly close to global authoritarianism. This is an existential risk all on its own.
Such an entity would have to:
Limit access to space (to avoid self-sustaining expansion faster than everyone else).
Limit/monitor the uses of AI and AI training.
Possess a credible monopoly on violence.
Surveil enough to prevent surprise attacks and overthrow attempts.
Considering the risk that such a system is corrupted or misaligned or permanent, I feel better about a future that emphasized freedom and acceleration of defensive technologies.
(I could be convinced that the “multiple night watchmen overseeing each other” is viable. Rather than oversee each other, it might be better to give them completely separate jurisdictions. Federalism and free movement allows people to choose night watchmen that suit their needs. Risk of war between jurisdictions is low since they both have watchmen. Some watchmen may allow AI’s to develop and entities to leak into space, but this laxity a feature to avoid global totalitarianism.)
I feel like people are dismissing this study out of hand without updating appropriately. If there’s at least a chance that this result replicates, that should shift our opinions somewhat.
First, a few reasons why the common counterarguments aren’t strong enough to dismiss the study:
I’ve been seeing arguments against this result based on vibes or claims that the next generation of LLM’s will overturn this result. But that is directly contradicted by the results of this study, people’s feelings are poor indicators of actual productivity.
On Cursor experience, I think Joel Becker had a reasonable response here. Essentially, many of the coders had tried cursor, had some experience with it, and had a lot of experience using LLM’s for programming. Is the learning curve really so steep that we shouldn’t see them improve over the many tasks? See image below. Perhaps the fact that these programmers don’t use it and saw little improvement is a sign that Cursor isn’t very helpful.
While this is a challenging environment for LLM coding tools, this is the sort of environment I want to see improvement in for AI to have a transformative impact on coding. Accelerating experienced devs is where a lot of the value of automating coding will come from.
That aside, how should we change our opinions with regard to the study?
Getting AI to be useful in a particular domain is tricky, you have to actually run tests and establish good practices.
Anecdotes about needing discipline to stay on task with coding tools and the cursor learning curve suggest that AI adoption has frictions and requires tacit knowledge to use.
Coding is one of the cleanest, most data-rich, most LLM-developer-supported domains. As of yet, AI automation is not a slam dunk, even here. Every other domain will require its own iteration, testing, and practice to see a benefit.
If this holds, the points above slow AI diffusion, particularly when used as a tool for humans. Modelling the impact of current and near-future AI’s should take this into account.
Yeah I think sleep probably serves other roles, I just don’t see why those roles require 7 hours of sleep rather than say 5 hours.
I do agree that basic research is what will actually get sleep need reduction therapies to work at scale. I’m hoping that citizen science and discussion of the topic will encourage more work on this.
That’s a great post! It did more to popularize the idea than I ever could. I’ve been thinking about this for a while and my first writing on the topic was in 2021. I’m going to refrain from linking to it because I’m planning on depreciating that blog soon though.
Re storage and handling: this part was tricky, we opted to dissolve in sterile water (not saline) and froze the batch after mixing. So doses were only exposed to room temperature for ~minutes. We also used a C-terminal NH2 peptide that is less susceptible to degradation.
There’s still no guarantee that peptide is stable under these conditions and we don’t have a good way to check. This is a big reason why we thought there was a ~60% chance of a null on this trial (and the next one too perhaps). But hope springs eternal!