> I think over-thinking is only solved by thinking more, not less.

These errors compound exponentially: the probability of not selecting wrong token for N steps is 1-(1-P_e)^N.

The shorter "thinking" is, the less is the probability of it going astray.

My answer of "how to judge?" question is the question "how easy is it to implement new unforeseen functionality with the code under scrutiny?"

  >  Most programmers suck at programming so badly that LLM/AI production IS better than 90+% (possibly 99%+).

  > AI hasn't observably reduced the quality of software I use everyday in a way that is meaningfully separable from normal incidents in the past.

Average duration for AWS outages [2] was 1.5 hours per outage, 38 hours total. Most recent AWS outage in 2026 [1] downed AWS for 13 hours, a third of 38 hours spanning an year before, and was caused by AWS LLM coding tool.

[1] https://www.theguardian.com/technology/2026/feb/20/amazon-cl...

[2] https://www.cherryservers.com/blog/cloud-outages

  > I have been writing a library called "assume" where you can specify a type signature, give it a prompt, and it generates a function on the fly in the background with Claude Code, so you still write some code, but whenever you need a function you "assume" that such a function exists.

[1] https://mail.haskell.org/pipermail/haskell/2005-December/017...

And this is why I boldly compare current LLM craze to the much less hyped craze of strong type systems. I was a part of that strong type system discussion, advocating for them. ;)

  > If you make the prompts specific enough and provide tests that it has to run before it passes, then it should be fairly close to deterministic.

  > I assure you that's already happening. A lot.

Google with Softbank invested about $230M into QC last year. Microsoft, IBM and Google have spent on QC $15B combined, through all of the time they researched it. $15B spent in 20 years, less than $1B per year, by three companies.

Google spent upwards of $150B last year in datacenters.

This may tell us something about how close we are to a working quantum computing.

While there are no companies with $1.5 trillions (4*$380B) of net revenue, the difference is that Anthropic is cash net-negative, has more than 4 people in staff (none of them are hungry artists like PG) and hardware use spendings, I think, are astronomical. They are cash net-negative because of hardware needed to train models.

There should be more than one company able to offer good purchase terms to Anthropic's owners.

I also think that Anthropic, just like OpenAI and most of other LLM companies and companies' departments, ride "test set leakage," hoping general public and investors do not understand. Their models do not generalize well, being unable to generate working code in Haskell [1] at the very least.

[1] https://haskellforall.com/2026/03/a-sufficiently-detailed-sp...

PG's Viaweb had an awful code as a liability. Anthropic's Claude Code has an awful implementation (code) and produces awful code, with more liability than code written by human.

The situation there is akin to Viaweb - Viaweb also rode hype wave and code situation was awful as well (see PG's stories about fixing bugs during customer's issue reproduction theater).

What did Viaweb's buyer do? They rewrote thing in C++.

If history rhymes, then buyer of Anthropic would do something close to "rewrite it in C++" to the current Claude Code implementation.

[1] https://en.wikipedia.org/wiki/Sauna

Hammam's temperatures are around 40-50 degrees Celsius and humidity is close to 100%.

These are very different conditions, with very different body response.

  > You recognize that you haven't really needed strong mathematical (or coding) skills to create models for some time.

[1] https://arxiv.org/abs/2304.15004

  > No one has ever made a purchasing decision based on how good your code is.

  > How can the same model predict egg prices in Italy, and global inflation in a reliable way?

So, predict sign (branch predictors in modern CPUs also use neural networks of sorts), exponent (most probably it changes slowly) and then predict mantissa using Benford's law.

  > A system that is not even Turing complete is extremely limited.

  > The Delta Cycle logic is actually quite similar to functional reactive programming. It separates how a value changes from when a process responds to that change.

  type S a = [a]

  register :: a -> S a -> S a
  register a0 as = a0:as

  -- combinational logic can be represented as typical pure
  -- functions and then glued into "circuits" with register's
  -- and map/zip/unzip functions.

Eurisko demonstrated superhuman abilities to play strategy games in early 1980-th, and even used strategies from VLSI place-and-route task in planning fleet placement in games. This is knowledge transfer between tasks.

[1] https://www.vice.com/en/article/a-human-amateur-beat-a-top-g...

  > This is meaningless.

  [1] https://en.wikipedia.org/wiki/Constructivism_(philosophy_of_mathematics)#Constructive_mathematics

The paper [1] referenced in your link follows the lagacy of the paper on the HIGGS dataset, and does not operate with quantities like accuracy and/or perplexity. HIGGS dataset paper provided area under ROC, from which one had to approximate accuracy. I used accuracy from the ADMM paper [2] to compare my results with. As I checked later, area under ROC in [1] mostly agrees with [2] SGD training results on HIGGS.

  [1] https://arxiv.org/pdf/2505.19689
  [2] https://proceedings.mlr.press/v48/taylor16.pdf

So, my questions are: 1) what perplexity should I target when dealing with "mc-flavtag-ttbar-small" dataset? And 2) what is the split of train/validate/test ratio there?