Michael Griffiths

Every line of dialogue in that movie is important to the plot. It's amazing. Fabulous movie.

Shape constrained splines are nice too, as a way to introduce theory-based constraints.

This is a fairly technical but highly relevant paper on how we can model complex systems at various levels of detail without losing causal content. Think gas: instead of tracking every molecule, we can focus on big-picture properties like temperature and pressure. www.auai.org/uai2017/proc...

Damn right.

From the outside, it just looks like Llama missed a trick focusing on MoE instead of reasoning. A bunch of people used that as evidence that the strategy was bad and convinced Z to blow things up. So far the announced moves have been an incoherent land grab at Z's expense.

Yeah, the degrowth stuff needs to be killed thoroughly and without mercy. No.

While true, these tech advances generally benefit society. The Luddites lost their jobs and their kids made less money for generations - but cheaper textiles was a huge benefit to many others.

My experience is that it felt like 4 for the first year.

I should! I just finished a P-spec for a system that looked something like this -- generate P spec from code, use that to find errors, turn error traces into tests in project language. You need to be a little careful but it's very fun.

Validity In Psychology Research: Types & Examples In psychology research, validity refers to the extent to which a test or measurement tool accurately measures what it's intended to measure. It ensures that the research findings are genuine and not d...

I took a number of methods/stats courses in undergrad, and the *only* discipline that covered this was, oddly, psychology. There, they spent a lot of time on validity (e.g. www.simplypsychology.org/validity.html) and then introduced stats as part of it.

It's great. Sonnet 3.7/4 is also good. I've been using TLA+ for months now to check my own code, and to explain system dynamics in e.g. bug reports. Growing adoption should increase your target market, not decrease it.

Great to see this (cacm.acm.org/practice/sys...) from @marcbrooker.bsky.social

My experience is that LLMs make using TLA+ much easier. For instance, I just wrote up a bug report last week that outlined the old/buggy behavior with a TLA+ spec. It made the dynamics much clearer.

Reminds me of "in the long run, we're all dead "

(Not always true, sometimes it's much harder)

Yeah, that's true. Verification is often much easier however.

This is one of the great uses of LLMs, though. Turns into (often) a 2-minute job

Mmmm, they don't. Even if they lose money on average per query, that's dominated by long outlier conversations.

I agree it's a big deal. I grew up in Santa Rosa and recall the fire that burned down >7% of the city housing stock. Drive through neighbors and see melted cars and lone brick chimneys. It's sad to see the same kind of thing in LA

I do think licensing limits available staff and drives compensation up, so provider cost seems like a part of it. The important takeaway is the incentive structure of insurance conglomerates.

Interesting take that capped profit structure pushed insurance companies into self-dealing and reduced cost control incentive to maximize profit

Reinforcement Learning: An Overview This manuscript gives a big-picture, up-to-date overview of the field of (deep) reinforcement learning and sequential decision making, covering value-based RL, policy-gradient methods, model-based met...

An updated intro to reinforcement learning by Kevin Murphy: arxiv.org/abs/2412.05265! Like their books, it covers a lot and is quite up to date with modern approaches. It also is pretty unique in coverage, I don't think a lot of this is synthesized anywhere else yet

How much money do we think means testing this is going to save? Like, how many ebike purchases would there have been for people making >3x poverty level?

Yeah. And insurance company policy changes get attention, but not the fraudulent billing or practices that lead to insurance companies cracking down.

Nice to see alternatives to TLA+ pop up - this case study of Fizzbee makes it looks like a nice contender for certain use cases.

Depends if you want to make the declaration of the variable more obvious. You can do

result <-
df |> f() |> g() |> h()

or

result <- (
df |> f() |> g() |> h()
)

... when you want people to focus on the `result` variable vs. the chain logic.

Quite - the lack of empathy to his death is something that bothers me.

Each transaction inside DSQL runs in a customized Postgres engine inside a Firecracker MicroVM, dedicated to your database. When you connect to DSQL, we make sure there are enough of these MicroVMs to serve your load, and scale up dynamically if needed. We add MicroVMs in the AZs and regions your connections are coming from, keeping your SQL query processor engine as close to your client as possible to optimize for latency6.

Neat! brooker.co.za/blog/2024/12...

Posting some evergreens for the new crowd. Did you now you can differentiate RANSAC?

If you fix the # of iterations, RANSAC is an argmax over hypotheses. You turn the inlier count into your policy for hypothesis selection, and train with policy gradient (DSAC, CVPR17).

github.com/vislearn/DSA...

Yes! My quality of life foes way down when FRED doesn't have something and I have to try to extract it from Eurostat or the OECD. Or even BLS for things FRED doesn't pick up

That feels like a timing thing? eg you built your workflow and became comfortable with it prior to LLMs, so of course there is no gap.