llm_trw

This comes up every time someone wants to give the death penalty for rape.

If the punishment for rape is harsher than the punishment for murder than anyone committing it may as well remove the evidence by using a blender.

We have human teachers for much the same reasons.

Are humans just other humans hiding in boxes pretending to play chess?

Other than sugar none of those are spices used for food.

And even sugar became much less commonly used to in Europe.

The wildest thing about European colonialism is that it was kicked off to get more spices for European food. Then in the worlds most bizarre twist European food gave up all spices a few centuries later and colonies had to invent something else to sell back to the colonial centre.

That's the third major use of the printing press.

The first was for catholic indulgences.

The second was for religious propaganda during the European wars of religion.

Ironically Linux with KDE is very good for being both pixel perfect, responsive and enjoyable to work with.

What a time to be alive.

It's not a product. It's business core competency in the ml space.

Data is king. Even when a new better model comes along a high quality dataset is still just as valuable.

Paying top performers above market rates to do nothing but data labelling is a moat that just keeps getting deeper.

Linus isn't a renewable resource unfortunately.

Longer than Germany at the rate things are going in Germany.

Britain had two parliaments in the home island - Scotlands and Englands.

The only people who can possibly think the British empire was centralised are those who have never opened a history book about it.

I suggest you look up how much raising children costs today vs the 1980s in both money and time.

Two kids today are a much bigger status symbol than a 5 million dollar house.

California has half the population of Germany and twice the GPD.

We can't pretend the US is some dinky country in Europe that can be ruled the same way.

What works for a country the size of a New York neighbourhood doesn't work for a country the size of a continent.

It's not terribly interesting. There have always been two strains of thought for the Democrats in the US. Once has been that we must move the whole country in the right direction kicking and screaming - this has been the ideology that has been in ascendency since Kennedy. The other that we must be allowed to be as progressive as possible without interference from the Federal government. That's not been in vouge since the end of the new deal.

I'm completely in the second camp and would move back to the US if it were possible to get it adopted at the national level.

Decentralise government so a single election doesn't destroy everything.

In short: states rights.

This is what gets me.

Homer Simpson was meant to be a fat idiot barely hanging on.

In the 40 years since the first episode he's gone to being average weight and fitness.

Economically he's now wildly successful for being able to own a house on a single income with three kids. There are people making over $800,000 today who can't afford that.

The rest of the world is just 30 years behind the Us.

The Netherlands today are fatter than Missisipi was in 1985.

The 40% of Americans who are obese don't work at Google.

Is there any evidence for general impulse control?

I'm seeing papers claiming both increased and decreased impulse control.

AI is not good enough yet for anything requiring deep reasoning, mission-critical work, error detection at a human-expert level, or handling unpredictable edge cases.

Ai is better than humans at all those things. It's not good at those things when the context it needs to look over is more than a few thousand tokens.

Rejoice programmer, for your inability to write modular code saved your job.

That's your problem. You keep the fixes for a rainy day when production is down and the business is losing $10m an hour.

Yeah boss I can fix it, but how much is it worth to you since this isn't in my job description.

He nearly got assassinated by the incompetence of the same people telling him to help Ukraine.

Mozilla has been pink washing itself for a decade.

This has paid off incredibly well for them since anyone pointing out that it's primary purpose was to enrich the CEO while letting her band of merry women support their pet causes was called everything from a fascist to an incel.

Till today the most blatantly farcical situation was that Mozilla funded women who code boot-camps while firing all the women who code inside Mozilla.

I guess people are finally waking up when Mozilla pulled an Ubuntu and decided that everything you do through their software is something they own actually.

They also flag and harass anyone who points out the grift.

A dataset of labelled answers that are hallucinations and not hallucinations are published based on the benchmark as part of a paper.

People _seriously_ underestimate just how much stuff is online and how much impact it can have on training.

You can if you do it on radio.

You don't pay for the physical medium, you pay for the right to play the song to other people.

Porn is not illegal, either are the Rambo movies.

In general yes, bench mark pollution is a big problem and why only dynamic benchmarks matter.

You can email me if you're interested.

This was paid work until recently but the start up pivoted to something else and now I'm wondering what to do with code that takes a few thousand dollars to run every time and is of questionable copyright ownership.

The logical calculus is easier to execute than any of the techniques of mathematics itself, yet only in 1934 did Gerhard Gentzen set it out in a natural way. Even now, mathematics students are expected to learn complicated (epsilon, delta)-proois in analysis with no help in understanding the logical structure of the arguments. Examiners fully deserve the garbage that they get in return.

With an opening like that how can I not read the rest of the book?

This actually touches on some work I've been doing the past week - arbitrary term rewriting to test how reliable a given LLM is when it needs to reason about symbolic manipulation. Everything is dynamically generated to avoid the usual problem of memorising every validation set every llm seems to suffer from.