red75prime

7 ms·

61.
by red75prime 2mo ago
The latest LLMs (except Qwen and DeepSeek) are MLLMs (multimodal language models). Unless you count RGB values as symbols, they are dealing with more than symbols. Yes, there are functional gaps between MLLMs and humans. Their long-term mem
62.
by red75prime 2mo ago
I guess it will take quite a while before we reverse engineer the human brain to find all the optimizations and shortcuts that evolution has used to make the human brain reach intellectual maturity in just about 20 years.
63.
by red75prime 2mo ago
My issue is that I can't predict what you'll find surprising, if an LLM (with a harness) would be able to do it. If you are able to check that the results of an LLM are satisfactory, it means that the results are checkable. Then,
64.
by red75prime 2mo ago
Do you experience dogs as lots of action potentials traveling along axons and lots of neurons doing their thing in your brain? I don't know how it gets from the physical processes or the information processing to our first-hand experie
65.
by red75prime 2mo ago
I hope such persons will be met with a spray of indelible paint.
66.
by red75prime 2mo ago
Waymo has never disclosed this information.
67.
by red75prime 2mo ago
> But once you get used to it and start to identify the patterns the brain quickly labels most AI-generated content as blank space. I guess the majority of people do low-effort generation that doesn't perturb a default style of a ne
68.
by red75prime 2mo ago
> However LLMs deal entirely in symbols. They use tokens as input/output encoding. They do 99.9999% of processing in a high-dimensional latent space.
69.
by red75prime 2mo ago
"Pattern prediction" is a very broad stroke. What's something AI can never do that would astonish you if it did?
70.
by red75prime 2mo ago
I think it has very little to do with reasoning and much more with psychological inertia or pattern matching.
71.
by red75prime 2mo ago
> Just like "thinking" implies chain of thought, but you'll frequently get those reasoning traces and then a 180 in the final message. Be me, think extensively about an exam question, do 180 because the most probable optio
72.
by red75prime 2mo ago
...because it is illegal to copy copyrighted material. 70 years later they might do it.
73.
by red75prime 2mo ago
I'm not a big fan of arguments like "it's not the real [human quality], it's [mechanistic explanation]." They lack a part: "because the [human quality] allows us to do X, Y, Z, which is impossible with [this me
74.
by red75prime 2mo ago
> if it could, you arguing with it isn't going to make it "learn" or act differently. Are you talking about a specific harness that doesn't have context retention mechanisms? For example, ChatGPT with disabled memory
75.
by red75prime 2mo ago
When I was taking an MIT AI course (in ancient pre-LLM times), an autonomous agent was defined as a system that perceives its environment and acts on it (we were focusing on reward-expectation-maximizing agents, but it's not that impor
76.
by red75prime 2mo ago
I reacted to a specific claim about a specific article. I don't have a statistical summary of all articles (even an intuitive one). You seem to have such a summary. Do AI-unrelated articles cause "we need to react to this urgently
77.
by red75prime 2mo ago
The article doesn't meet a bug report standard. How would you feel about a bug report that doesn't mention which software were used, which versions were used, but it extensively describes all the harms that were caused by the lost
78.
by red75prime 2mo ago
It's easy to bring down the hammer of scope insensitivity by vividly describing a single person's struggle. The question is how AI affects error rates.
79.
by red75prime 2mo ago
Yep. I was born in the USSR, and my mother still occasionally marvels at one doodad or another. The last one was a reusable bowl cover.
80.
by red75prime 2mo ago
It's not a proof. You check that the mathematical ideas expressed in the claimed statement are the same as the mathematical ideas expressed by the Lean repository.
81.
by red75prime 2mo ago
> On the Self-Verification Limitations of Large Language Models on Reasoning and Planning Tasks An empirical paper from 2024 that doesn't give a principled reason why LLMs will always be bad at self-verification.
82.
by red75prime 2mo ago
A subdivision of people who know every obscenity in the world would be chuckleworthy, but I guess they use a validation program.
83.
by red75prime 2mo ago
> If the training data doesn't include lots of text of people being asked questions and saying "I don't know", then it's unlikely to respond "I don't know" There's a problem with your reasonin
84.
by red75prime 2mo ago
> I'm saying that any reasonable expected rate of breakthroughs is not enough to maintain exponential R&D effort "Any reasonable expected rate" does a lot of work here. For example, I haven't expected LLM contribu
85.
by red75prime 2mo ago
Unfortunately, we aren't there yet. Whyalla Steelworks was intended to produce "green steel" using hydrogen. The project began around 2022 and was cancelled in 2025 (timelines weren't met, financial troubles). Port Bonyt
86.
by red75prime 2mo ago
Generalist humanoid robots that are able to operate in unstructured environments weren't an option (they still aren't an option for the majority of operations). The humanity had no ways of building them. Anyway, it's true tha
87.
by red75prime 2mo ago
It requires large solar overcapacity to compensate for seasonal intermittency. Dedicated seasonal energy storage might be better economically.
88.
by red75prime 2mo ago
Predictions that are based on "no breakthroughs, no significant capability changes with scaling this year" might fail too.
89.
by red75prime 2mo ago
I guess it's the largest by its useful load.
90.
by red75prime 2mo ago
4) Algorithmic improvements are either relatively easy to find if you already know the system can do better, or they don’t provide an edge that can’t be overcome by increasing training compute.

More ›