So, let’s see. Since my last summary nothing has been changed much, except the US stock market, Nasdaq100 and S&P500, in particular, are printing ATHs every single day and MU and DELL, and even INTC and AMD – you can see it yourself.
On the other side Anthropic emerged as a true leader in the armrace [to the bottom] since their code slop generator is qualitatively better than Gemini – the closes competitor. Everything else is just bullshit (both ChatGPT and Grok are, indeed, merely Infotainment chatbots).
What exactly Anthropic did so right is unclear, definitely it has nothing to do neither with its delusional CEO anthropomorphic abstract babbling about matrix multiplications becoming “sentient”, nor with the “scientific” papers about high-dimensional manyfolds and what they imagine the actual representation is (it is in the actual algorithms being used, not in the papers being published).
Perhaps, the winning streak was in both augmenting/decorating the code slop with abundant verbiage slop, both in comments and in accompanied “documents” and reports/summaries, and in feeding this very slop back as new training examples and in the Reinforcement Learning post-training (to “myelinate” or reinforce the most used pathways, so to speak).
The facts, however, remain unchanged – since the generator (an inference engine) is still a conditional probability based or a Bayesian one, a generated slop, however “intelligent”, consistent or “correct” it may appear to a human observer, is still a cognitive illusion, in principle. It still operates at the level of abstract tokens, and the generation or inference process is still inherently sequential, meaning that once bullshit (sorry, a slight hallucination) has been introduced into the context (of what has been generated so far) it would determine, and, indeed, condition the rest of all the potential outcomes, just like turning onto a wrong road on a road fork.
This explains perfectly the manifestations like “how many week days have the letter ’d’ in them” (just try it!).
And yet, the code which even the free-tier crappy web interface produces is simply “hands down”. This is the actual power of the computational bruteforce – the size of the models and the quality of the training data (I guess they were smart-enough to threat different programming languages as different human language, which is kinda obvious not to mix).
All the stupid memes about being a “prompt engineer” (puke.jpg) have some truth in them – if you know how to constrain and restrict the slop generator and, like a sculptor, know from experience what comes first and what has to be done much later, you could get code slop of such a quality, which can be attained only after months of systematic, focused, attention to details based, principle-guided effort, but in just a few minutes. I have seen this happen almost every day.
Yes, Gemini deletes its own comments it placed there a few minutes ago, deletes the whole blocks of code instead of just fixings single typo, switches back and forth between large chunks of wrong over-“engineered” slop, but it eventually compiles and runs under the strictest lints (which is the only right way to do it).
Now what? I guess we begin to appreciate when IBM said a 10 likes of fully debugged and documented code per week is okay, and paid something like $60k in Brooks times.
The steep or the slope of the race to the bottom the coding LLMs would force us into is simply mind-blowing. And believe me, they will commit and ship the crappiest imperative spaghetti slop which barely compiles and passes some tests.
The only thing I know for sure is that the world will never be the same again.
And yes, the business model of charging per token on a subscription basis (the dear darling of any MBA, second only to charging per breath as per meal or per place to sleep packed as tight as a stack of coffins) would never generate the hundreds of billions spent on hardware, construction contracts, connectivity and what not. Cognitive illusions, however clever as sophisticated (the combined efffort of the brightest minds of humanity in many fields, after all) would not print enough money to pay back the loans, especially with the ongoing rat race to the bottom, which the whole programming industry naturally (ship whatever crap doesn’t crash every minute as fast as possible) became.
Yesterday I have read somewhere “they told us to study for 15 years and then replaced us with an LLM subscription”. Also imagine that the only thing you know and familiar with is some webshit framework and how to navigate throughthe accompanying fucking mess – React of whatever.
I have studied my whole life. Here. There.