AI CEO: ‘Stock Crash Could Stop AI Progress’, Llama 4 Anti-climax + ‘Superintelligence in 2027’ …
The latest on Llama 4, and whether it signals a slowdown in AI, or solid progress. Plus, a deep dive on that viral prediction of superintelligence by 2027, and Dario Amodei’s cautionary words on what could stop AI progress in its tracks. o3 news, and more, as well.
Weights & Biases:
DeepSeek Doc:
AI Insiders ($9!):
Chapters:
00:00 – Introduction
00:47 – Stock Crash
02:28 – Llama 4
10:55 – o3 News
11:59 – OpenAI non-profit?
13:13 – AI 2027
Llama 4 Release:
Dario Amodei Comments:
Knowledge Cut-off:
Aider Polyglot:
Gemini 1.5:
Fiction-LiveBench:
OpenAI Valuation:
OpenAI Cybersecurity:
Deep research System Card:
AI 2027:
METR Paper:
OpenAI non-profit:
NYT Piece:
Kokotajlo predictions 2021:
Non-hype Newsletter:
Podcast:
Good morning from Canberra P. 🇦🇺
the goat has returned
Benchmarks are helpful, but what truly matters is how these models perform under real-world conditions; outside the lab, with messy data and unpredictable input.
See end of the video !
the harder benchmarks are the ones closest to reality, like simplebench with it’s trick questions
The problem with AI right now is not messy data and unpredictable input, we mastered that already. We are using ai with the new quantum chips, AI is the only thing that can understand the Quantum chips response.They use Ai to translate the Quantum chip responses to human readable data. The real major problem is energy consumption. Because they very complex mathematical algorithms used to decode the messy and unpredictable inputs uses gpu and cpu which requires a lot of energy to stay cool and running
The reason it seems like ai has no real world use is because most human don’t even know how to use human intelligence in real world conditions. If you look deeper in reality you can see AI is already making its impact. Since along time ago. The thing is people think AI just got invented. What people think is AI is regenerative conversation ability. That’s actually not the AI intelligence core. The core is a very complex mathematical computations. Mathematical computations made human civilizations through out history, AI has always been here.
@gemstone7818 And let’s be real, it’s still a basic approximation. There is so much tacit knowledge tied up in engineering teams that is not possible to “write down” – and yet is context that the AI lacks.
Disappointed in Llama 4. Any chance you could cover non-LLM architectures, like DINO or JEPA?
second this. jepa looks really interesting
LMArena rankings seem to be a complete joke these days
You’re conflating a recession with what Dario Amodai said “a belief that the [AI] technology will not move forward”.
Many reasons for a stock crash or lack of capitalization, which can be causal of the technology not moving forward also
12:35
“or whether AGI is even feasible in the next three to five years”
I love how longer than 5 years until AGI is the slow case scenario
I guess EU will just use the Chinese models instead. It is bananas how they regulate everything here. Even the shape of bananas.
Well good. You cant have data floating out to foreign companies.
Either its running local or its useless.
As someone that follow AI from pretty far (you and bicloud pretty much) I tried to read the LLAMA4 release note before this video and yeah, this is why I subscribe here, I got something right and compared to deepseek v3.1 and got some of my own tough but I clearly do not have the depth and knowledge you have and even less able to articulate it as well so I always feel like I lean new thing and your take are pretty cold and grounded and in a world of hype that is very VERY much appreciated.
Many thanks for the work put in, it must bot be a lite workload seeing how fast thing go
When will we see a model that stays updated with the present moment with a high sample rate while continuously learning new concepts?
Dario Amodei did specify the loss of trust in the technology itself rather than the loss of capital, i don’t think people are specifically losing trust of ai over this
But his broader point was that a crash would starve capital, he didn’t anticipate tariffs qiuld cause that, not loss of confidence in AI
@@aiexplained-official sure but if confidence in the technology hasn’t waned then capital is most likely going to get restored, and it is also uncertain how long these tariffs are actually going to last
Thank you for covering the AI-2027 report/scenario (not a paper btw). If you ever get the chance to play their tabletop exercise, would you consider livestreaming it? or at least, making a video about your experience with it? (you need at least 10-15 people and 3-5h btw lol)
The prediction reads more like science fiction that actual prediction.
Just remember, the mid-level engineers prediction is coming from the same guy who thought we’d all be watching this in the metaverse.
Great as always.
When I start working on a new project. I open a new AI window and will converse with AI throughout the day, on my project, in that same window.
Does the “nearly infinite” 10 million context thingy mean that my conversation can be longer and the AI will remember better… OR is that 10 million only relative to the prompt input?
12:57 Why is OpenAI no longer the dominant player in the race to AGI? Who are the dominant players?
Looking at the exponential progress of AI *in general* since 2022, it wouldn’t shock me if they got it right for the prediction in 2027. Because if fully automated RSI is achieved, it would create ASI almost instantly. I think we can all agree on that. Afterward, it is not unreasonable to think that such a model could create a perfect (or close to it) simulation of the real world. And that would bridge any sim2real gaps. I’m not completely confident that will happen; my original prediction for this was 2029, but considering the sheer progress in the last 6 months alone, I’m starting to think that the 2027 prediction is plausible (so long as something catastrophic such as a Taiwan war or 1929-style depression doesn’t happen).
Amazingly interesting! Thank you so much.
Using MyLMArena chrome extension to automatically track my LMArena votes and turn them into my own ELO leaderboard. And yeah, on that Llama 4 is right down on the list, even below Phi-4. No idea what other LMArena users are doing rating its responses so high! Gemini 2.5 has the most Elo on my leaderboard. And I now use it in Cursor, dethroning Claude Sonnet for me for the first time!