Philipp Cannons @pcannons

@bryan_johnson Bryan, am I drinking 3 Monster Energy Zero Ultra every morning for the bugs in my brain?

@Migrat0ry @menhguin Exactly, 6.5h-7.0h is the tracker equivalent in numerous studies for 8h TIB. Hitting >7.5h on tracker consistently is indicative of an underlying condition.

@zachmeyer @swyx My desert. My Arrakis. My Dune.

I guess the belief is these neurons don't have enough structure to support consciousness. But since we don't really know how to measure it, it's not a certainty. And it becomes an "increasing concern" as more complex biological organoids are built...

I hope this thing is not conscious: youtube.com/watch?v=yRV8fS… The monitor for Oxygen, Carbon Dioxide, and Temperature to the machine feels very weird. And the fact that the neurons die after 6 months.

Would be great if we figured out how to measure when there's consciousness so we're not accidentally torturing a bunch of things.

Dr. Alex Wissner-Gross @alexwg

5 months ago

x.com/i/article/2029…

772 2K 11K 11.3M 8K

@willccbb Interesting, that was Gemini 2.5 and now 3 Pro for me. I feel Opus 4.5 is just not as knowledgeable outside of engineering tasks to chat with. Feels both have clean pre-training but Gemini just has more broadly spanning data.

@karpathy Once LLMs have culture, I wonder what kind of conspiracy theories they'll be into, and who the Joe Rogan equivalent + Pyramid obsession will be

@TheAhmadOsman Ya, makes sense, figured they might be a bit dated now. I didn’t think about speed. No build currently, was going to build a new rig.

Ozempic for mental health, incredible! I have a small passion project from 2016 with about 11,000 lifetime users around weight loss. About 100 people lost more than 30 lbs, and 5 people achieved extreme fitness. I only worked on it for a few months over the years, but I’ve gotten to talk to and learn from hundreds of people using it to understand the problem better. The reason I never turned it into a full-blown company/product is that I realized weight loss - and I suspect many more things, like finances - is fundamentally a mental health issue. If you can’t let the mind overcome boredom, stress, and anxiety, a meal planner, a calorie tracker, etc., are just bandages and won’t meaningfully help. It’s been great seeing mental health become less taboo over the last decade, and seeing effective techniques - such as meditation/gratitude (separated from spirituality) and self-affirmations (separated from horoscopes) - become more mainstream. This could be a great addition for those in need. An argument can be made: "Why are drugs needed for this? Shouldn’t natural methods just work?" That would be great, but very unnatural things are happening in the first place. For example, first there’s a well-meaning innovation: the Interstate Highway System. It’s used for bad things: serial killers. Then it’s addressed by technology much later: DNA testing. I believe Ozempic is a similar counteragent to cheap, addictive food. And I imagine this could also be in the toolkit for modern problems of the mind caused by unnatural things like social media and 24/7 global news. Super early. Hopefully this doesn’t also become a "well-meaning innovation" and is safe. But promising!

Dillan DiNardo @DillanDiNardo

10 months ago

1/ It feels surreal to announce completion of the first human trial in the development of our neurotech platform for designing mental states, from the molecular level. Human experience is now programmable. A🧵on the sequel to psychedelics & the first new "emotion in a bottle."

170 411 3K 365K 2K

@maxhodak_ This feels absolutely groundbreaking. Kind of surprised it got such a weak reception. Tough to do product demos I guess.

@michael_chomsky Yep, Zotero is great for this. I also like to pair it with MyMind

@dexhorthy Look forward to reading it @dexhorthy My post, granted it's evolved a lot since: philippcannons.com/100x-ai-coding…

If you're curious what the "big bet" is right now at frontier AI labs, here's a quick non-technical summary: The models have effectively been trained on all available human knowledge (internet/books/videos/pictures) for a while now. But we kept making the models bigger to make them better at "predicting the next token". From GPT-1 to GPT-4 it worked incredibly well but has plateaued going from GPT-4 to GPT-4.5. Along the way, in GPT-3, a technique of incorporating human feedback to show the token prediction engine how to "chat" with us was discovered and that's how ChatGPT was born. There has since been a realization that we might be able to use that same technique to teach the models how to think. (o1, o3, GPT-5 thinking). What's "thinking"? The model generates a bunch of text hidden from you to show its work. Then it says after which ChatGPT starts showing you the text it generates. We've even hooked up a computer program to run based on these "thoughts". e.g. If the model thinks "Use Tool, Google: What's the weather today?" a computer program will go and Google the weather and return the results. The model can then incorporate the result of that search into its thoughts. That's what the "reasoning models" are doing, that's why they take so long. It's been shown the models aren't doing "real reasoning" though, in the human sense. So the idea right now is to get lots of examples of good reasoning and behavior to hopefully teach it how to. So once again a "hypothesis of scaling" but now through behavioral data (coding/computer work) and thinking traces. That data is much harder to get your hands on so there's a bit of a gold rush to figure out how to get lots of it at scale. It's also an open question: if you train a model on a massive and diverse enough dataset of "problem solving" from countless different environments (math, code, science, logic puzzles), will it eventually stop memorizing specific solutions and start learning the underlying, abstract reasoning operators (e.g., decomposition, deduction, pattern matching)? If so, can that finally create Artificial General Intelligence (AGI), intelligence that can do anything a human can do? That's why all the labs are creating "coding agents", buying coding companies (Windsurf), and buying digital "environments" and behavioral/thinking traces. Those are treasure troves of complex digitized reasoning and verifiable outcomes. e.g. You can observe all the steps it took an engineer to solve a problem and how they went about it. So it's an open question. And some top AI researchers think we need a new approach and this won't work. But for now, Occam's razor, we'll see if this approach will create AGI or it'll be on to the next thing.

@algekalipso Next do the same poll but for an AI consciousness

@haydendevs hayden never learned what an MCP was. “I was too poor,” he said. “And then I was too rich.”

@Hesamation Haha, I just made this meme and the timeline immediately shows me your meme x.com/pcannons/statu…

Philipp Cannons @pcannons

10 months ago

heartbreak part 7

1 0 3 828 0

heartbreak part 7