AI Breaks Its Silence: OpenAI’s ‘Next 12 Days’, Genie 2, and a Word of Caution
Calmest before the storm? Whatever analogy you want to use, things had gotten quiet toward the end of 2024. But then tonight we got Genie 2, and a series of scheduled announcements from OpenAI…
Assembly AI Speech to Text:
Sora is soon here, and o1, but I dive deeper into what it all means and whether reliability is on a path to being solved, ft: two recent papers. Plus Kling Motion Brush, Simple Bench QwQ update and much more.
AI Insiders:
Chapters:
00:43 – OpenAI 12 Days, Sora Turbo, o1
03:06 – Genie 2
08:26 – Jensen Huang and Altman Hallucination Predictions
09:45 – Bag of Heuristics Paper
11:40 – Procedural Knowledge Paper
13:02 – AssemblyAI Universal 2
13:45 – SimpleBench QwQ and Chinese Models
14:42 – Kling Motion Brush
Genie 2:
Jim Cramer:
Give Us Full o1:
Verge Scoop:
O1 Learning to Reason Benchmarks:
SIMA AI:
Genie Paper:
My Video on Genie:
Oasis Minecraft:
LLMs Procedural Knowledge Paper:
Bag of Heuristics Paper:
Jensen Huang Hallucinations:
DeepSeek Interview:
Kling Motion Brush:
Tim Rocktaschel Book:
The 8 Most Controversial Terms in AI:
Non-hype Newsletter:
Podcast:
Many people expense AI Insiders for work. Feel free to use the Template in the 'About Section' of my Patreon.
The OpenAI announcement AND a new AI Explained video? Merry Christmas everyone! 🎄
The best part about the next 12 days, will be your 12 videos breaking it down.
God that might be a bit much! But I will be following everything scrupulously, don’t worry
@@aiexplained-official we expect 12! No more no less.
I want at least 3 videos. 4 is enough.
@@countofst.germain6417 Best he can give is one 12 minute video or twelve 1 minute videos.
Maybe in the format of those one-minute shorts.
You are so fast! Best AI news channel imo
Definitely!
If we solve hallucinations I am kicking off my personal “we are now in the future” project. I have spent the last year developing a procedural immersive fantasy world simulator _with the entire design portfolio_ formatted as instructions for agentic AI. I have roughly 400 pages of instruction from power balance formulas, to world history instantiation, magic systems, UI/UX, an LoD system for simulation granularity and information retention, on and on. it’s been bounced off Claude from start to finish at each step so I know interpretation is solid.
Such a thing _will_ exist. I will make certain of it.
nice! Would love to test it out when it becomes playable 🙂
me too!
cool
exciting! I’d definitely want to try it out once it comes out. So far, the only real game that is AI-driven is AI Roguelite, as far as I’m aware (I bought it but don’t necessarily want to try it out just yet, as I’m waiting for 3rd party API key support since the dev told me he is considering adding it and it might come in a few weeks). AI Dungeon doesn’t count, it’s not really a game. Some could argue AI Roguelite isn’t either (yet). That’s one of the things I’m most excited about.
But also, on the other hand, I don’t actually think solving hallucinations is a trivial problem and it might never occur without architecture change, so your caveat about that definitely tempers my excitement. Would love to believe, though, lol
@@maciejbala477 indeed, as the video explains, hallucinations are in some sense necessary for model creativity. if you want something to generalize, it needs to know how to fantasize — this is not avoidable since it does not know the latent space you expect it to generalize to, it must guess it.
I turned the quiet period into a blueprint. By mapping every workflow (medical deparment in big pharma) , I’ve created a roadmap for the AI Agents that’s already knocking at our door.
Literally like an hour ago I was scrolling through youtube and thought ‘I wish AI Explained would upload again’
Awesome video as always by the way
Thanks N8
lies 😂
The hallucination problem is this – it will be a long time before a model can be the fail safe for a process that has to judge physics or maths & get it right 1 shot every time. It puts a lot of uses which an Ai would be perfect for out of the picture for the foreseeable future anyway.
these are definably guessing machines, their errors don’t vanish, like if it learned to preform actual mathematics, they only shrink.
@@anywallsocket The same is true for the best human mathematicians!
@@fabim.3167 haha very true! but not the same for typical functional programming — which is what people colloquially associate ai with, and hence all their surprise when it gets stuff wrong.
thats exactly the problem with LLMs they need a graph backbone – middle layer – for solid representation of functions, symbolic abstracions and the world. the creativity is well needed for creating that middle layer and for creative work but not for most tasks
probably they could be interfaced with a game engine, which they could learn to control to generate situations, but invariably the engine would compute the resulting physics — then you can feed-back on that and cound get some mildly effective self-optimizations.
Did anyone notice the defects in aerial shot of the neighborhood? There are some driveways that don’t actually connect to the street, along with other unrealistic aspects.
Yeah, the roofs of the houses in the middle blend into the road etc. It’s impressive but it’s a lot like every other AI image generator – seems great until you _really_ look at it and then there are almost always weird flaws.
(which is great in one sense – we can still _pretty much_ tell an AI image from reality, maybe except in mostly neutral categories like landscapes etc.)
@@anonymes2884 yeah it’s just HD slop
Thanks again for your great content. I love the fact that you can avoid click bait and still give us compelling headlines
Thanks Jason
good insights from those papers that temper the expectations, looking forward to the full review of those papers
Being under a million subs this far in is insane. Your ‘spidey sense’ with the nuance of ai progress is unbelievably precise.
Thanks Keat
Philip Pharker?
Yeah, but later you can be all hipster about being here when his channel was small
Glad you mentioned China’s push in AI towards the end of your video. Despite working with comparatively limited hardware, they’ve managed to shrink what could’ve been a multi-year “moat” to almost nothing. It’ll be fascinating to see how this plays out in the future!
@@Isabelle-w7i I mean Llama is at GPT3 levels and is open source so it’s not hard to see how the world’s largest electronics manufacturer country could catch up
All humans use the bag of heuristics approach, it’s called thought. Even top physicists find counterintuitive physics in the world, and think in a collection of rules of thumb.
I was going to say, humans also hallucinate spherical cows. 🙂 Reductionism is a thing.
Also, I have the impression that the hallucination problem is more a matter of expression than of psychosis. We all imagine in our minds, then choose (hopefully) what to say. The models should be free to “hallucinate” inference-time tokens, but then choose to voice out loud only facts than it can confirm, or at least qualify those facts that it cannot.
I’m overly excited about all your coverage over these next 12 days. Hopefully OpenAI surprise us with really exciting and cool announcements.
learning a ‘bag of heuristics’ rather than explicit maths is how i skipped through my undergraduate degree lol
Nice
Like all Engineers hahahahah
by far the best AI youtube channel. No unnecessary hype, no baseless rumors, no creepy AI art for thumbnails, just pure objective analysis.
That is so funny. I can hardly understand him he is talking with the mouthfull of marbles, no doubt🙄
The creepy AI art for thumbnails is god awful
Are you serious or such naive?
@@marpleka I’m especially serious about the AI art thumbnails from other AI news channels
I think the hallucination problem drastically drops once reliable, high-quality agent and function calling are out and accessible through an LLM. We just need something like a beefy 01 that has access to tons of tools that it calls reliably, and as long as those tools work properly, it’d be a huge leap forward. Fingers crossed that in the next 12 days OpenAI has something agent related out
I appreciate that you’re one of the few ai YouTubers that actually admits there hasn’t been big news recently and doesn’t overhype same stuff just for views.
The solution to having one apparent model that does well at hallucinating creative work and also reliably does physics is just higher level mixture of experts. The user converses with a model that is specialised in interfacing with the user, and it has access to other specialised expert models as tools. That way the prose model can level up independently from the physics model and the biology model, and so on. It comes back to that jagged frontier.