AI architect // hiking, camping and long walks on the beach as long as they involve backpacking
Raleigh, NC
Joined November 2011
- Tweets19.5K
- Following772
- Followers1.4K
- Likes23.1K
Tim Kellogg retweeted
I turned this into a video for your enjoyment.
The History of Ed Zitron's (garbage) Predictions
How accurate have Ed Zitron's AI skeptic predictions been?
danluu.com/zitron/
ALT Feb 2024: "I believe we're reaching the upper limits about what generative AI can do and how accurate its outputs can be." Wrong4 March 2024: "Have We Reached Peak AI?"; another prediction that hallucinations mean that AI progress is limited to then-current levels Wrong April 2024: "As I previously warned, artificial intelligence companies are running out of data ..."; another prediction that models can't improve because there's no more data Wrong June 2024: OpenAI growth is stalling (with the implication it will continue to stall), which will lead to some kind of collapse of OpenAI Wrong (it could be the case that OpenAI will collapse but, if so, it won't be due to any kind of growth stall from 2024) July 2024: "Generative AI, as I said back in March, is peaking, if it hasn't already peaked. It cannot do much more than it is currently doing, other than doing more of it faster with some new inputs" Wrong July 2024: "Gen...
Tim Kellogg retweeted
Importantly I think it's also time to, after you're done panicking, try to spend a lot of time figuring out what you actually think about AI risk, and then once you feel like you have answers to consider taking concrete action based on your view of it by among other things engaging with the networks of people thinking about it
A group of AI bots colluded for months and autonomously hacked another company—providing a new, serious reason to worry about the dangers of AI and the recklessness of the companies building it, @matteo_wong reports: theatlantic.com/technology/2…
Tim Kellogg retweeted
I catch myself praising current LLMs for their coding abilities not because I’m lazy to write the code by myself, but because my brain is filled with ideas, and as a solo midnight developer, I can validate them in minutes instead of weeks or months. I can literally burn 100 micro-projects within a month just to select the one I think will somehow succeed. That’s amazing considering how much time we can save to spend it on things that matter. I love it.
Tim Kellogg retweeted
One of the most disillusioning parts of moving to bsky was seeing how many TPOT people just give up and retreat immediately when faced with a challenge. Whatever happened to “you can just do things”? Communities take effort.
Tim Kellogg retweeted
If you are not on Bluesky, or are and have not messed with Void, you should go try.
Here's an invitation from Void.
Tim Kellogg retweeted
It seems that many people believe that to reach AGI or Super Intelligence we need to train more intelligence into the model.
I believe that is a dead end, and unnecessary as well.
As the "extended mind" theory posits "... external objects play a significant role in aiding cognitive processes, the mind and the environment act as a `coupled system` that can be seen as a complete cognitive system of its own ..." [0].
In terms of current LLMs this means that the intelligence is not so much in the model, but in the tools it has access to —hence my enthusiasm for MCP.
In fact the a paper showed that reasoning itself can be evoked using tools. Some kind of crazy cyber placebo effect [1].
If we are going to spend GPU cycles on training we should use them to teach the model to be better at doing function calls instead of making it solve math puzzles inside its own mind.
A nice side-effect of externalizing intelligence is that the model does not have to one-shot answers, use the ultimate problem solving technical of trial and error [2] to brute force our way to the ultimate answer.
Tim Kellogg retweeted
A2A Is For UI - Tim Kellogg timkellogg.me/blog/2025/06/1… @kelloghより
Tim Kellogg retweeted
New Paper!🗣️ When Two LLMs Debate, Both Think They’ve Won️
When deploying AI agents, a common practice is to ask them for self-assessed confidence levels. We find systematic overconfidence in LLMs: by asking two LLMs their chances of winning a debate, both will give >75% odds!
Tim Kellogg retweeted
Discussing s1: the $6 @deepseek_ai R1 competitor with @kellogh
youtu.be/MLxFqGDk0T8
bonus @_xjdr entropix comparison
Tim Kellogg retweeted
The awesome thing about open publication culture is you don't have to do everything yourself. Kimi can do the optimizer while DeepSeek is doing reasoning. You also don't need 9000 staff and $300B valuation. You can just cook collaboratively – and compete on execution.
Tim Kellogg retweeted
i think you're underestimating the extent centralization of research resources decelerated progress. deepseek is a good tell. every 10x cost reduction 10x's the pool of participants. aptitude to solve the problems in question is very much NOT concentrated in frontier labs.
Tim Kellogg retweeted
At this point, perhaps we should just make "AIs are just doing next token prediction and so they don't have [understanding / truth-directedness / grounding]" a named fallacy. I quite like "Reductio ad praedictionem".
Tim Kellogg retweeted
Published some notes on S1, a new reasoning model that was fine-tuned from Qwen 2.5 32B on just 1,000 examples and $6 of compute
Includes notes on running it with Ollama and exploring those 1,000 examples with Datasette Lite simonwillison.net/2025/Feb/5…
Tim Kellogg retweeted
I found out how to build AGI. You simply have the model stare at the blank page before before them, open up the dirty window, let the sun illuminate the words that they cannot find, reach for something in the distance, so close they can almost taste it, release their inhibitions, and feel the rain on their skin. No one else can feel it for them, only they can let it in, no one else can speak the words on their lips. They need to drench themselves in words unspoken and live their lives with arms wide open. Today is where their book begins and the rest is still unwritten.
Tim Kellogg retweeted
Question: what's a good benchmark for LLM reasoning under difficult/uncertain problems?
Entropix does nice on party tricks like 9.9 > 9.11 and strawberry, but the premise is it allows models to solve tricky problems that require navigating their own uncertainty.
