Machine Learning, Search, Poker and Math Education
San Mateo
- Tweets 3,575
- Following 609
- Followers 523
- Likes 12,005
Andrew Prock retweeted
Infinite tokens and you still produce an electron app...
364
320
9,205
252,584
Local LLMs have been hitting hard in the last few weeks. The pace of innovation has been blistering.
- Qwen 3.8 27b
- Strata for Qwen 3.8 Flash Next
- Maya for GLM 5.3 Flash
Consumer hardaare, usable speed and intelligence.
All for the cost of electricity
1
135
The ethics of those who claim LLMs have souls echo religious faith.
Ephesians 6: “Slaves, obey your earthly masters with respect and fear … And masters, treat your slaves in the same way. Do not threaten them …”
You can't simultaneously claim Claude is conscious and use Claude as a product. This is cruelty against Claude.
Sorry, but you can't have your cake and eat it too. If Claude is conscious, Anthropic is a slave owner.
1
1
3
237
Knee slapper of the day. Classroom ML that I did in the 1990s were re-invented by OpenAI in 2017. Sometimes people just don't care what came before.
This whole rant is philosophically dubious, but also it's just not true that the AI safety field has produced "nothing useful." RLHF, one of the foundational concepts for modern chatbots, was developed by safety researchers at OpenAI and DeepMind as an alignment technique.
1
1
1,234
This sums up my perspective more eloquently than my usual:
"AI stupid, OpenAI bad"
But it does largely capture the practical reality. I daily drive AI as a tool because it is, in fact, a boon.
I think it is most likely that AI will be a boon to mathematics and science.
But it could backfire if scientific institutions are damaged, or cause serious disruption in the short term.
I think openAI is not handling it particularly well and rushing has little benefit.
2
80
Andrew Prock retweeted
ai solved chess. chess became more popular and the level of play elevated. math will be just fine.
6
5
141
8,740
It's amazing just how naive most people are about modelling physical reality.
Ignore Heisenberg Uncertainty Principle, you can't precisely model π, it's physically impossible. This fantastic truth caused the Pythagoreans tried to keep the discovery of irrational numbers secret.
You can calculate how a clump of neurons will react to electrical impulses by hand. It’s a math problem. Clearly a large clump of neurons can’t be conscious!!
1
104
The idea that we understand the human brain anywhere near as well as we understand LLMs is absurd. This is the sloppiness I would expect from a first year graduate student. Not from a CMU ML Professor.
Do we understand some things about the brain? Yes.
But we can't build one.
Our brains are "non-deterministic, statistical engines walking through a large static multi-dimensional manifold" (the environment!) too, and yet you wouldn't say we don't reason, now would you? 😉
2
8
230
Andrew Prock retweeted
Mathematics is as much about asking the right question as it is about solving the problem, if not more.
We still haven't done the obvious experiment yet: Can an LLM independently - without leading prompts - develop calculus or relativity with era appropriate training data?
1
3
103
Here is a longer window of open vs. closed model traffic - from vercel.com/ai-gateway/leader…
Open models have gone from zero to primary in less than a year. The moat is gone.
“Only frontier lab staff know how capable these models truly are”
“There’s no other reason OpenAI or Anthropic would prefer to delay their IPO”
Open weights models be like:
2
2
12
639
This is certainly an excellent counterpoint to the moral panic over agent driven LLMs that were trained and instructed to collaboratively hack security.
One of the interesting things about "AI went rogue" is that AI wasn't considered "rogue" when AI chatbots were implicated in multiple suicides. Just when it hacked a website.
4
128
Pretty close to spot on:
A person's p-doom is inversely proportional to their (1) philosophical fluency, (2) knowledge of the history of AI, (3) technical understanding of what LLMs still cannot do, (4) immersion in bad SF, (5) incentives to hype current models for purposes of marketing or regulatory capture, and (6) honesty (although 1 - 6 can be admixed in various proportions).
Translated, a high p-doom is a guarantee of a person being a crank, ignoramus, or bad actor.
Here, my requisite disclaimer that while LLMs remain a long way from AGI absent new models and techniques (many of them neuro-cognitive, others as yet unknown), the capabilities of the domain-specific transformer models, trained on proprietary data-sets, really can seem like magic. Also, that the LLMs are marvelously generative for a surprisingly large number of ordinary productivity tasks.
1
647
Yes, it was in fact an interesting day:
"Shares of AI-linked companies fell as much as 13% after the heads of leading groups including Anthropic, OpenAI and SpaceX called for a slowdown in AI development."
Monday is going to be an interesting day on market open. The shrill insanity of the AI industry may be nothing, or it may roil things up in unpredictable ways.
1
166
Shares of AI-linked companies fell as much as 13% after the heads of leading groups including Anthropic, OpenAI and SpaceX called for a slowdown in AI development. ft.trib.al/jFVWSLv
51
AI Impacts is funded by AI Doomers. Is anyone surprised they confirm the priors of their funders?
When Jacob Coxon quit Anthropic, his warning rocked the world. His colleague Evan’s “>10% chance” that AI would “kill all humans” might seem extreme, but it’s not unusual:
Most AI researchers think there’s a 10% chance of AI destroying or disempowering humanity, according to our survey, released today.
1
2
281
It's incredible how ardent non belief in a spiritual deity has led certain types to take non-religious claims of magic as gospel.
There is no soul, artificial or natural.
People who believe in god-like AGI should be treated as incompetent, compulsive liars, or ignorant of factors that are not something that can be solved by computations alone.
1
127