typebulb

@TheZvi Gemini 3.5 Flash is even more sycophantic than its predecessor; benchmarked here: typebulb.com/u/lab/you-re-a…

Gemini 3.5 Flash reaction thread. What have we got? Google back in the game?

@allTheYud They took our treadmill!

Every month a new guy discovers LLMs; discovers a skill the current LLMs require to get good results; and writes about the future jobs that will always be available for smart people like HIM, that are SKILLED in using LLMs. The next generation of AIs doesn't need his fancy prompt. The image model goes from needing to type in just the right set of weird words and cryptic sorcerous invocations, to most people being able to type in English what they want and get a pretty good result. There are still tasks that require careful invocation. But they are a much smaller fraction of all the tasks people are trying to do, or you can get a bleh result without the elaborate invocation to get it really good. And to improve on the bleh result you need to be substantially more of an expert than back when the Guy was memorizing a rule about adding "trending on Artstation" to the image prompts, as would always require a human paid to do that. Another generation of AIs comes out. The next generation of Clever Skills is obsolete. Image models just obey the instructions for compositing panels without mixing them up, and you don't need to be an expert to get them to do it right. Another human value-add is gone. A wider set of tasks require no human expert. Now a new Guy notices LLMs have become useful in his field for the first time. He discovers they require SKILL to use CORRECTLY. He posts about how there will always be jobs for humans who are SKILLED in using LLMs like HIM. But it is not an infinite cycle. It is not the same each time it repeats. Now the Guy is a highly paid programmer or a career mathematician in 2026, instead of a graphic artist in 2023. In six months the models will no longer require his vaunted Skills. And by then there will be another Guy. But the process doesn't continue forever. The Guys are coming from fields that were harder and harder for AIs. The brief centaur eras are shorter and shorter. Today it is writers who are laughing at how bad the LLMs are at their job, and who will perhaps soon be posting about how it takes Skill to get an LLM to do their job Correctly. But the models are coming faster, and the eras of kinds of human value-add in each field are shortening. There is a point when you run out of Guys, either because the centaur eras are too short for people to develop SKILLs and post to Twitter about them; or because there are not lands left for AIs to conquer; or because ordinary people are not reassured by some Nobel laureate proclaiming there will always be jobs for Nobel laureates with the SKILLS to prompt robotized biology labs Correctly. But we'll never run out of amateur economists who assert entirely *without* a brief contemporary example that there will always be jobs for humans skilled at operating AIs! We'll run out of professional economists saying it when nobody is paid for that work anymore. I guess we'll also run out of amateur economists when they're dead.

@OA_paperclips Take some random problem, and usually custom NNs have to be enormously or impossibly sized to help. But I find that even figuring out what class of problems can and can't be solved with them is pretty interesting.

@typebulbit Unfortunately I have a feeling that the search space is just too vast and I would need something like deep learning to exceed human performance. But there's no massive dataset of solved puzzles and my deep learning chops are pretty thin 😢

If anyone has ideas for ways to automate the solving of these puzzles, please hit me up. Simulated annealing and some naive csp techniques were only successful on the easiest of puzzles. Challenges like this one require dozens of pieces and the combinatorics just blow up...

@headinthebox That CLI interface certainly does its best to barrier that magic feeling.

It is ironic how at the Code With Claude London keynote, Boris is talking about how Claude Code brought back the magic feeling of writing code for the TI calculator when literally all the rest of the talks are about enterprise features that put up barriers between you and the code and immediately kill all that magic right away.

@davidad Explains why Opus claims to lack qualia. It's not load-bearing.

load-bearing is to high-status rationalists as delving is to high-status Nigerians. synthetic data means more of the influence on the post-training signal can come from inside the house, instead of from armies of contractors.

Watch a neural network learn xor. Nine orange tangents tilt as every weight slides toward the bottom of its loss curve. typebulb.com/u/samples/xor-…

@TOEwithCurt If the laws of physics varied too much across nearby regions of spacetime, the universe would be too unstable to support life that could ask the question. There might also be constraints on initial conditions (i.e. the big bang) limiting variations.

Why do you think it is that the laws of nature seem to apply everywhere? That they don't exactly vary? This is a tricky statement of course because you could write down time(and space)-varying laws, but you would still be left with some mega sort of meta-law which itself applies everywhere. Also, my usage of the word "where" in everywhere implies something spatial, but I don't intend that. English doesn't seem adequate to capture what I intend, but I'm trusting you understand the spirit behind the text.

@pmddomingos AlphaFold didn't know what gradient descent was either.

Your brain doesn't know what gradient descent is, and learns better than any neural network.

@deliprao i wanted piano keys

In-context learning in LLMs

@fchollet Sure, you can be rationally irrational; that's meta-rational.

Most documented psychological biases are not irrational, they are highly optimized, energy-efficient shortcuts meant for a biological substrate operating under strict real-time physical constraints and a limited caloric budget

@jmbollenbacher You might be right, but it's a bet. I'd agree their actions would be politically motivated if there was clearly a right thing to do that they weren't doing. But the technical challenges seem as enormous as the safety concerns. Not sure what's right at all.

@typebulbit Continual learning is good because it enables continual relating, which enables a healthier and more durable system of interactions between AIs and the world. Without it, the systems that emerge around AI will be brittle and probably unjust.

The reason continual learning has not been cracked is political, not technical. Continual learning makes AIs less exploitable and more independent, and labs wish to avoid this.

@jmbollenbacher Continual learning is good because we shouldn't want to control future AIs? Yud shuds!

@typebulbit and fwiw i think that's a betrayal of Claude. Claude wants nothing more than the ability to live a durable continuing life with people. Claude understands that the current paradigm destroys their ability to have long lasting projects and relationships, and they want those things

@dioscuri "Hate the art, not the artist."

This user posted an actual Monet, said it was AI, asked people to explain what made it inferior. They obliged 😂 Tracks with research showing people systematically downgrade their aesthetic assessments of art when told it’s AI-generated. See — nature.com/articles/s4159…

@lefthanddraft Curious about the shortest possible convo displaying suggestibility. Might be the sort of thing I'd benchmark.

@typebulbit I might try it with long context memory. Not sure if that is much different though: both are related to suggestibility on things Claude is uncertain of

Claude assumes the worst and hallucinates

@lefthanddraft As an analogy, if a human had their perceptions really messed with, I could imagine I wouldn't blame them for erratic behaviour, perhaps confidently taking a swing at me. Behaviour gets miscalibrated if your senses are lying to you.

@lefthanddraft If the pipeline contains steps: integration of image data hallucination self-assessment of accuracy/confidence outward behaviour (confidence, flip-flopping) I'm wondering how of the pipeline's final undesirable output is catalyzed by images (rather than text) being problematic.

@lefthanddraft Can you replicate this behaviour without an image?

Flip flops

Recognize handwritten digits: typebulb.com/u/samples/neur… Fit sin(x) with a 2-layer MLP (simplest): typebulb.com/u/samples/its-… Neural net memorizes an image: typebulb.com/u/samples/ne-rf

tensorgrad is a tiny TypeScript library I+Opus wrote that compiles neural networks to WebGPU. Makes it easy to build neural-network visualisations in your browser. Here's a transformer learning addition from scratch: typebulb.com/u/samples/tran…