Time To Think

I don’t remember what it’s like to think with more than 60 seconds of context.

This may sound like an absurd exaggeration, but I’ve tried to make the number well-calibrated to my current reality.

Keeping the 3 sentences of context above in full mental view is a struggle. Not a metaphorical struggle, but a physically painful stabbing pain at the front.

So, my ability to write has degraded. In 2021, it was effortless to recall a day’s worth of activities for a CTF, describing events with an implicit planned scope of writing end-to-end.

In 2025, I cannot. Honestly, I have no idea what I was doing even 1 hour ago, or why I started writing this. If someone told me straight to my face that I was a test experiment for the limits of human simulation at low context lengths, I would take it at face value.

All of the above is immature melodrama. Physical CT/MRI scans gave a clean bill of health; reasonable people do not dupe themselves into believing in psychosomatic bullshit.

Let’s talk about AI cognition speed – but only from a personal reflective PoV, I’m in no position to talk about Deep Learning from a technical point of view these days.

In the far long distant past of early 2023, GPT-4 was released. Back then, I looked at the apparent progress (90th percentile exams, wow!) and would very seriously joke about AGI coming by 2025 (to anyone who would listen back then anyway). It didn’t matter that it was slow as molasses or objectively worse at python than me – it was obvious the next jump would fix it.

And then, for some reason, that jump did not appear in public. For nearly two years, text-only ChatGPT users could not observe anything in public beyond increasing inference-optimization related brain damage.

So, much like in the intro, I duped myself – even if I never explicitly said it, my actions throughout 23Q4-24Q3 had naturally bent themselves to fit the profile of someone who disregarded the possibility of fast superintelligence altogether. Doing the most pointless things for the most narrow-minded reasons.

Now, in 2025, my original expectations have become simple fact. Inference-time scaling was the jump, and there’s nowhere left to run.

Not disregarding the fact that all language models, by their nature, have a vast diverse knowledge space many times larger than any individual human, it was still possible to identify model shortcomings for topics I had specialized in. Even if tiny, the set difference between e.g. o1 and me was still nonempty.

o3, on the other hand, is a plain intellectual superset. I do not know any useful information it does not know, I cannot perform any AI-friendly knowledge work it cannot, and of course it does it at blazing 100x speed, 100x less cost, and 100x more focus.

So, tough luck, sucks to suck, it’s time to move on and graduate together with the 99% of first worlders that have already accepted their incompetence under the machines, right?

No, of course not, AI is just a tool, you like tools, tools can be leveraged, leveraged to build new products, to build something people want–

If I could do any of that I would not be wallowing into a text box. Forget building a product, I am lucky if I am able to think hard enough to mash in a question for superintelligence to handle for me on any given day, let alone implement anything remotely resembling a plan that requires more than 4 hours of goal-directed behavior to succeed.

…

I don’t remember what I was going to continue writing for this section. Probably some pseudo-economic nonsense about how the degredation of the economic value of intelligence via non-interventionist tools is uniquely damaging to the lazy+smart archetype.

In spite of its incomplete state, I will simply press the Upload File button, because I am aware that the last 9 post drafts I wrote were all also left to decay in limbo after I forgot about them.

Actually, their titles were kind of cool, let me share them:

1
2
3
4
5
6
7
8
9

11548 Mar 27  2024 'State of FP8 Training.md'
4654 May 26  2024 'The future of real-time translation.md'
6076 Jun 17  2024 '2026 Predictions.md'
15K Jul 18  2024 'The MLA Problem.md'
6.5K Jul 26  2024 'Spacebyte++ - A Failed Replication.md'
6.8K Oct 18 10:38 'Accurate Description of Tensor Parallelism.md'
2.7K Oct 19 00:31 'Recreational Homelessness.md'
33K Oct 24 18:58 '(Tutorial) Blackbox Hyperparam Tuning for Transformer Pretraining.md'
27223 Jan 15 07:18 2024.md

See? Real. Ignore that an LLM would easily create a similar mock-up. Or maybe you have the intelligence to confidently determine its veracity.