Barry Kelly

Do they subjectively feel like they've overcome LLM limitations to you?

It's great that they've been fine-tuned on a bunch of synthetic symbolic problems. Benchmarks go up.

But I still need to clear the context whenever a bad take or mistaken assumption gets in.

2/6
This is the point I have been making again and again over the years. The global economy is a closed system, and it must balance. This means that domestic imbalances created by countries that control their external accounts must...

(At this point I still prefer to use AI for any expensive-ish (say, over three digits) purchase, which tips the scales towards feeding the data monster, i.e. reviews, influencers etc., getting people talking.)

Sure they're mostly aligned, but the misalignment has an important inefficiency: auction bidding directly siphons off buyside margin to sellside + broker - that's money that doesn't stay in the consumer's pocket. So even if I see an ad, I rather re-search with that vendor than click through.

Search ads are basically a parallel search engine using ad targeting criteria and auctions to determine ranking, instead of an analysis of the user's value function, which is far too expensive. Until AI, that is.

The numbers are somewhat skewed, because the incentives to improve avenues for sellers to buy placement are at odds with finding the best response for the user. If you spend more effort on the former, it looks like people prefer the former.

Persuasion goes both ways and facts can be cherrypicked, and are, heavily, by biased publications everywhere on the spectrum. Facts and evidence are, alas, no protection from malpersuasion.

(If it's not clear, corrupt authoritarian politics are a bad thing, and should not be supported wherever you are on the political spectrum. But the public en masse often doesn't realise what it's buying when those guys sell fantasies.)

The Trump administration's enemy isn't countries, it's non-authoritarian politics. This is why Europe is a bigger threat than China or Russia.

It's also why Russia attacked Ukraine after Euromaidan. Liberal (as in, not based on authoritarian crony power networks) politics is the threat.

That doesn't capture the vibe of slop to me. Slop is all texture and nothing more.

Slop is all surface but hollow on the inside. Aesthetic polish, but nothing below the surface. Low meaning. Information content entirely in signifiers but nothing signified. A walking, talking zombie with dead eyes.

It's the other way around. LLMs are adept at content-free PR speak, so PR needs to adapt to avoid being mistaken for LLM slop. IMO.

This is not that different from the old Cluetrain Manifesto. Authentic voices don't sound like LLMs.

A high ratio of linguistic polish to meat always smells like LLM to me now. Like, show me details that are falsifiable. High powered verbs, adjectives and adverbs but lacking specifics: who, where, when, what, how, why. Paint me a picture of process that fixes things and improves proactively.

All action words without specificity. "Comprehensive" steps. Secured & revoked (how many?), blocked (from where?), engaged (who?), performed review (what was found?), additional controls (what controls?).

It reads literally like a sample "best of breed" response. No evidence of engineering.

I take a train every week, between Zurich and Hannover. Between 60 and 110 EUR a leg 1st class depending on how far in advance booked. Food and drink to your seat but it's a mediocre choice. Either way it's far better than driving or flying. Only private flights would compete.

It happens quite a lot actually. Most Starbuck clones, for quite some time, do not offer black coffee. The best you can get is an Americano. It's the menu that irritates though, not the barista.

Have you ever gone into a coffee shop, wanting black coffee, and been faced with a menu without that on it? Where the closest thing is a watered down bitter espresso labelled Americano?

Maybe just me.

Akshually, I have, far more than once, been forced to drink Americano when I wanted black filter coffee. And it's a watered down bitter espresso and not very nice!

But what about paying people to wait on you?

Stuff only goes so far, and it's relative. The hedonic treadmill is real; my new PC in 2025 delights me no more (less probably) than my new PC did in 1995.

Getting people to do your bidding, on the other hand...

There's a problem with LLM problem solving that I'm seeing pretty consistently now.

LLMs, when they encounter an error, formulate a hypothesis about what's wrong.

Then, rather than validate the hypothesis, they immediately try to implement a solution based on the hypothesis.

Check the hypothesis!

It generates a backlash in some people because it situates moral culpability not in racists and their racist acts, but in having a skin color. If you have white skin, you are starting from a morally compromised position; i.e. you're a bad person because of your skin, is what it means to some people.

They can sell them to me? I bought my rtx 6000 pro Blackwell for a tidy sum but I'm open to buying more...

Is this news? TPUs have been around for ages.

I don't think he's measuring anything to do with popularity. I think Polanski is a nutjob economically, just a much less offensive nutjob than Farage, and if he gets anywhere he'll do a different kind of damage. Neither Farage nor Polanski can tackle the root causes IMO.

I have no strong feelings one way or the other about Cluely - maybe mildly disapproving of the cheating thing the CEO did but someone was going to do that eventually - but the zoning thing is kind of dumb. Beds under stairs isn't good, but mixed use is good actually.

And all the data. Don't forget all the data.

Because a bunch of hardware will be released into the market from companies that failed to make it in research, and that hardware can be used for inference. The marginal cost is dominated by electricity, and wholesale electricity is trending cheaper, esp. with solar. 2/2

When you look at batch throughput, the technical bottlenecks considering MoE models, and power use of hardware involved, there's no reason to believe the subsidy is there for commodity inference.

I know you want to believe otherwise, but a collapse will make inference even cheaper. Why? 1/

You are mistaken. Training, research and salaries dominate costs of the big cos. The marginal cost is not zero, but it's a fractions of a cent for a short query and response. Third party providers for open weight models have no incentive to sell at a loss.

Why be limited to what you need rather than what you want? Hobby room, library, cinema, guest bedroom, home gym etc.

Ironically, the best way to deal with terrorism is by treating it as criminality. Because military approaches oppress civilians and risk backfiring, motivating support in the wrong direction.