morelandjs

Man, can only imagine this is more bad news for delivery drivers.

I think what people fail to appreciate, is that past a certain size, organizations and the people within them get so tightly wrapped up in working a certain way that it becomes nearly impossible to fix without rebuilding the whole thing from the ground up.

If you want to see if AI is paying off, look at new startups that build AI-native versus those that don't, controlling for the type of industry/product. This is a difficult causal inference comparison to make, but it's the one that I find most meaningful and interesting.

I like how you have to be an accredited investor to get involved with riskier investment instruments but you can be a total moron with no money and lose whatever you have left betting on sports and prediction markets.

Btw it’s not just those two, add iGaming into the mix which is cell phone gambling glammed up in Candy Crush form. Leeches, all of them.

In many cases, you can partition data at rest (eg parquet) on the user access key, then load it once into an agent container at runtime. In this way, one authentication pass provides access to all data in a sandbox, versus repeatedly authenticating against MCP calls. It also just reduces the connective tissue between the harness and the data. There are obviously drawbacks in that you aren’t leveraging real time APIs, but it works well for most business use cases.

Was ready to write something snarky because this is essentially RAG, but I think the author is getting at some subtle details which are seemingly important.

- memory systems are a specific type of knowledge base where you generate all the documents. You might as well generate them to be less than your embedding token limit to obviate the need for chunking.

- embedding models are getting better and are no longer just semantic averaging.

- small models are getting dirt cheap, making parallel reads cost manageable

What they describe is sort of the simplest architecture that takes advantage of these observations. I believe them when they say it works well.

I do suspect though that things like keyword lookup will completely fail if every memory is just a vector. Hence why something like Typesense hybrid search can still be useful.

Worth mentioning that controlling for confounders is hard, a reasonable prior is the null effect, and any time a researcher makes a mistake it produces a publication claiming a novel effect. You observe a set of publications concentrated with false positives as a form of selection bias. I wonder how many of the causal studies are reproduced by independent teams using independent datasets.

Plausible scenario. Individuals predisposed to Alzheimer’s experience different mental sharpness from birth and this makes them enjoy taxi driving less (more intense than long-haul trucking), and so they pursue taxi driving as a career at lower rates. Under this scenario, driving has no effect, it just induces a selection bias.

Maybe confidence is a loaded term, but LLMs do have a sense of procedural uncertainty. Couldn’t you ask it to respond y/n in response to a question and look at the next token sampling probabilities?

Unsolicited endorsement, but I bought a “Brick” device recently, and I can’t recommend it enough for cutting back on social media. Infinite scroll really destroys your ability to concentrate.

Would love some kind of randomized A-B test where CEOs are replaced with AI and the outcome measured. Would not at all be surprised if it’s far better than the bottom quartile or more.

More practically, I think you could give a board of directors access to an LLM that sees all levels of company operations and let them talk to that chatbot as a stand-in for the CEO on a 24/7 basis.

Immediate page abandonment for me. I would appreciate more pledges of “this is entirely human writing, my own thoughts, imperfect as they may be”.

Using a better tech stack also provides opportunities to simplify.

Man, I hate that tech has just become money. Content like this perpetuates it.

I published in Nature Physics and the copy-editing process was quite embarrassing, to the point where we had to repeatedly nag them to stop them from making the manuscript presentation worse.

To be clear, I’m not talking about subjective style issues, I mean conforming to their own spec and avoiding careless bugs.

All remaining work fell on the backs of the physics referees. I’m not sure what value Springer provided from an editorial standpoint. It was disappointing to say the least after all that hard work.

Agree on article frustrations. Perhaps a better explanation, skills are just disk-cached prompts conditioned on verified success. The conditioned on verified success part might seem inconsequential, but it’s the whole thing that gives skills their value. Also the fact that their loading can be scoped to a certain calling context.