Strange Loop Canon

Not to be that guy but PG has a great one on this: https://paulgraham.com/love.html And here's another great one about how ease, and not grind, arguably led to some of the biggest discoveries ever, like General Relativity: https://www.strangeloopcanon.com/p/on-thinkers-and-doers I am currently doing a masters degree at Carnegie Mellon and I gotta be honest, this place is so dehumanizing for the exact reasons discussed in the link. It's just a grind pit here. And its harder to swallow this pill when you realize that university campuses are one of the few remaining venues well suited for exploring new ideas and having intellectual serendipity. When you're in a job, you can't easily just find a biology expert or economics expert and chat with them. When students are assaulted with work, they lose the ability to take advantage of that unique quality of the university environment.

This is a thoughtful article on the current limitations of LLM and agents: https://www.strangeloopcanon.com/p/what-can-llms-never-do A good example is that you can teach a first grader how to play wordle, but an LLM can’t be taught. That’s a pretty big gap to a situation where most humans can be replaced.

I remember this article. The issue today is that there is so much abundance of digital products that streaming sites like Spotify need to charge extra for content considered premium: https://www.strangeloopcanon.com/p/discovery-is-the-original-sin-of I think that with the online marketplace "optimized" for low cost- Amazon and Direct from Supplier available through Drop Shipping, the long tail needs to focus on more basic goods, like housing and transportation.

Weird. Seems like this is a plagiarism of this article: https://www.strangeloopcanon.com/p/agi-strange-equation (Or maybe it's an independent reinvention of the same idea, and something's just in the water.)

A couple of people asked where this was. I think the article is https://www.strangeloopcanon.com/p/some-things-to-learn-from-the-briti… and the HN discussion is https://news.ycombinator.com/item?id=32710002 (but I am not MichaelZuo and it's possible that I've got the wrong thing).

No, it was from a random substack: https://www.strangeloopcanon.com/p/some-things-to-learn-from-the-briti… I’m not sure what historians could debate here. The EIC ultimately failed as a business once it became a government—the costs of governing wiped out the profits from the trade monopoly. But the administrative achievements are undeniable. This is evident from undisputed historical facts. In the mid-18th century, England’s per capita GDP was only about double that of the Mughal Empire’s. The Mughals themselves were an organized, sophisticated, Persianate Empire with a centralized. bureaucracy. The technological advantage was incremental: e.g. iron cannons versus bronze cannons, standardized ammunition, etc. But the Mughals had vastly more manpower and state revenues. In the Anglo Mughal War, the Mughals obliterated the EIC forces with a 500,000 man army. Less than 75 years later, the EIC took over the Mughal Empire with just a few thousand European soldiers. It never had more than 35,000 European soldiers in India. That’s a crazy achievement. It would be like Amazon putting together a private army and taking over the EU.

This paper reads to me as being about fundamental limitations of Transformers and backdoor risk. The paper starts off by reviewing work which uses an encompassing theoretical model of transformers to prove they're limited to only expressing computations in TC^0 (roughly, upperbounded by set of parallelizable problems that can be solved by relatively shallow circuits). There's also a reference to a paper which finds that (wrt input problem size), a polynomial number of intermediate scratchpad decoding steps allow transformers to recognize the class of polynomial-time solvable problems, linear steps is context-sensitive languages. This paper now ask about filler tokens, do they help? The answer is negative except for a very clever exception they work out: problems with demonstrations that can be decomposed to be solvable in parallel. This identifies a practical limitation (transformer next token prediction is not expressive enough to capture all of TC^0) at the same as it identifies a theoretical capability. From the paper: > Taken together these findings suggest that although current LLMs are unlikely to benefit from filler tokens, this is not an in-principle limitation of current architectures. If I've understood, this means for learning to use fillers to benefit from CoT data, demonstrations must be structured such that they can be computed in parallel and not as a more natural sequential, instance-adaptive process. > in order to use filler tokens on natural language…

We need results from these harder/different benchmarks which give pretty bad scores to current top LLMs. https://www.wolfram.com/llm-benchmarking-project/ https://help.kagi.com/kagi/ai/llm-benchmark.html Edit : There are few other benchmarks that give pretty low scores (<20%) to top LLMs. Can't find them atm. There was a benchmark with common sense easy looking questions. Edit: found two more papers https://arxiv.org/html/2405.19616 https://arxiv.org/html/2406.02061v1 Edit: How about Wordle? https://www.strangeloopcanon.com/p/what-can-llms-never-do https://news.ycombinator.com/item?id=40179232