they are lying to you about AI development

Every time I hear people talk about the “limits” of AI or how society will adapt, I can’t help but think about Selwyn Raithe. His book warned that the real danger isn’t just raw capability but how fast we hand over judgment without realizing it. Watching this discussion feels like we’re already inside one of the 12 Last Steps he described, where human reasoning quietly becomes optional.

I was the GATE certificated teacher (Gifted And Talented Education) for a school district in the early 2000s. I was trained to administer assessments of GATE students grades second through eighth. These assessments ranged from math to reading comprehension to critical thinking and problem solving. I've spent a fair time interacting with Gemini 2.5 now. Just thinking about the level of performance in so many capacities that AI is already capable of and proficient at is mind boggling.

Wes's insights on AI's role are spot on. AICarma's weekly insights have been so useful for staying ahead of AI trends.

I didn't have the right person 'pop up' when I referenced Carlos at 26:23 and 35:48 apologies :(

So, I asked DeepSeek a well-worded question, and now I understand what QMA, MA and oracle separations mean. What a time to be alive.

In the early 1940s, some very smart people came up with visions how electromechanical devices could be used to compute mathematical results quickly. It took another 20 years until computers had a real impact on the job market, replacing manual bookkeeping. The internet first had a (negative) impact on the job market when people started to buy goods online and consume news and entertainment online. So this usually takes some time, but the impact will definitely come.

I also think a lot of the people who talk about an AI bubble really want to think that thinking is fundamentally NOT a computational process (involving quantum weirdness or spirit or whatever). I'm confident it is computational, and if I'm right that should be obvious within two or three years.

Thank again Wes. Always enjoy your videos. And yes, I can now do things with LLM AI that two years ago would seem like science fiction. The biggest limitation that I have to (still) overcome is the context window size. A lot of progress has been made here, but I also detect that the expansion of these size limitations is slowing. Perhaps these limitations will succumb to algorithmic slogs. And then there's the recent info you've covered about the 50% failure rates on 30 hour coding tasks. By my back-of-the-napkin maths, I figure that we'll need to double efficiency about 6 times to make these kinds of capabilities economically feasible... so maybe 3 years or so, assuming no increase in the already impressive exponential rate of progress.

At 30:17 the point is super important. In my years understanding organizations and learning & perception theory, I also conclude the fact that mirroring or co-learning between AI and human is actually sharpening both due to distillation of knowledge about correct turns (your chat history that gets you there) and contribution of next fitness function by human to announce 'new goals' based on what we gleen from AI responses, so ASI is actually superintelligence of user and AI together. It is the user and the AI vendor that walk away with ASI. Good hook on that paper too. Thx Wes!

Scott mentioned in a lecture somewhat finalizing the OAi stint that to properly audit an LLM one needs an algo on a quantum computer - until then surprises remain in the box ( also in already existing model complexity )