👍Check out the Playlist Physics Informed Machine Learning❗❗

Genius! Computational Bio-chemistry, for example, has been around for decades, but it typically uses a "brute force" approach: trying every possible combination. An "intelligent" version of that would be exponentially more efficient no doubt!

Agentic modularity also has some security benefits

I agree with your general idea that LLMs alone are not agentic systems. However, they seem to possess the ability to serve as a brain for an agentic system. Consider an application that has an agent orchestrator with the ability to call other agents or functions. All today’s agent frameworks use this concept.

Could it be that the wider variety of problems suggested in the unsupervised icl is just activating math related attention heads allowing it to solve the problems. where as QA might be less varied and cause more of a pigeonhole effect? I feel like this has been seen with some of deepminds RL work where less prescriptive performs better that prescriptive work.

This is what has been rattling around in my head for a long time. And I think it’s similar to what LeCun is saying. What we have is a model that produces language. Our language is not our intelligence. It is the communication of intelligence. It’s the final product. The wrapping around the true intelligence. The LLM should be used to encode my communication to the agent, then the intelligence (not LLM) goes to work, then the result is communicated back to me using an LLM. What we have at the moment is a very convincing and entertaining flea circus.

great video, perhaps Wolfram's TOE computational paradigm can be used and boosted with the development science-physical model

Nice I like it. I have been working on combining 31 model types in to one new model and your work is reinforcing and exciting. Thank you I hope more people see and understand your it's Fresh and cutting Edge.

😎🤖

My preferred architecture is open source and decentralized function calling language models running on personal devices such as smartphones that collaborate in building a shared world model that can include links to repositories of specialized models such as fluid dynamics, climate change, economics, and so on.

Isn't LLM more of high school geometry/trig that language?

The chat is gone, anyone have the chat exported?

I find it a little odd that they are going to try to bring out a product that is essentially already free to students and the general public everywhere (or $20 for monthly) and try to get universities to pay for it. The general features desired by researchers, doctoral and master degree seeking students are hallucination free excerpts, that accurately reflect the state of the art, with the ability to cite verifiable bibliographic references to ensure the quality of their papers. They "might" be trying to sell "teaching" services such as teaching a class to students. This would of course work pretty well to help students learn or to help grade papers to some extent. However, I can tell you that teachers are generally not crazy about having AI grade student papers, but rather would prefer they just give the students "advice" to "improve" their own work. AI already has an interesting alignment setup, but yea, how much integration should AI have ... and to be honest, anything that tries to "diagnose and treat" would be stepping over the line. However, if an AI is talking to a student that expresses some psychological issue that might result in harm to self or others, they might be legally responsible at some point. (It is a difficult issue). A student's Educational record is very private. In fact at the College level, we have FERPA which protects the student from releasing information. I don't think we should have AI tracking a student's records from early life thru college... the potential to do harm would be very great. I can imagine a sort of dystopian analysis of not just the "grade" that a student received but also a full analysis of a student. It would be a huge breach of privacy.

OpenAI will secure training rights on all the data that ebbs and flows through academia, and one day GPT-42 will be able to take any exam with 100% score, or score any exam with perfect accuracy or TEACH any class multimodally with 100% persuasion. Universities are dead, they just don't know it yet. It's a vampire deal.

3:00 it's really hard to think that universities would be early adopters.... but who knows

Probably an unmentioned sales proposition are features for managing academic honesty. I suspect that all student ChatGPT completions would be available to compare against any submitted assignments, which could inform an academic honesty metric in a GPT grading assistant for faculty. Stored completions across the entire institution is a huge value add, as opposed to having them spread out over multiple non-integrated products.

Interesting topic. Idea for future video. What are the main areas of the "planning" domain (A*, multi criteria decision making, analytic hierarchy process, etc). What planning algorithms are there that works well with LLMs? How to train LLMs to do planning?

wow, 4 minutes wait time for such a poorly written response? how far time has come. I run models on my gpu that answer all these in a few seconds at 150 tokens/second and at much, much higher quality output.

On the topic of attention and context, would love to see a video on Needle-in-a-hastack and multi-needle-in-a-haystack performance of these different kinds of context extension approaches.

What about release some code with your videos as well?

Totally agree that "Agentic" is either marketing hype, or we are are changing the definition of the word with a drastically lower bar to something approximating agency to function calling. And if AI agents are so drastically different to agents in society (i.e. human beings), should another name be given to it instead? What's really sad to me is real thought leaders like Andrew Ng also using the "A" word. But I guess I shouldn't worry about it if I don't understand.

Big misunderstanding The llm itself it's not agentic it's the orchestration of them through a specific automated system workflow that makes it agentic it's the process not the individual components that agentic

I think we need a fandom wiki to organise your content😅. Can you do a (have you done a) video on sbert fundamentals? How does it work? How is it different from bert? Should we use Roberta? Chat gpt uses all the right words, but doesn't help me understand.

You should check Maisa and their KPU (Knowledge processing unit). A novel approach that differed from RAG, function calling etc..

Yes, I hear our sales people throwing that term around........😅

This is another of your videos that I will have to load the transcript to gpt4o and ask it questions while I rewatch.

Agentic is future state of course there no agentic systems or frameworks yet it’s all still in beta and development, why are you trying to prove that all the agentic workflows are not agentic yet? You missing the point and why are all your videos sarcastic? Why don’t you build with langchain, langgraph, crewai, Autogroq and autogen and show us your brilliance by building early versions of what agentic workflows could look like? Why are your videos always so negative man? Who hurt you?

Years eh? Hehe. Have you heard of global workspace theory from Hofstadter? Agents+cognitive model+workspace=future

Such a wonderful presentation and understanding you gave in this video Would you like to see the next level from what you were showing if you would like to contact me I have a patented technology to show you

Sick vid brah

Scare tactics? So it is like WeChat. If you want to talk to a Chinese you must go through CCP spy and security channels. nice

I look forward to having a digital agent running in my smartphone that will be part of a global digital platform that will be able to have conversations with millions of people around the world at the same time, and merge the knowledge and sentiment expressed in those conversations into representations of the collective will of humanity. We will have collective human and digital intelligence.

You're awesome! So funny and informative! 😎🤖

Please tell me you did not soley let the LLM think for you on evaluating what agency means applied to LLMs. If it can make a random decision based on too much information than a model trainer can certainly to pre-decide what the model will output each time, then it's semi controllable agency by the time you give it task fulfilling function calling. It's no longer a 1:1 deterministic equation, it's a statistical calculation like our brains use. 🎉 Therefore LLM's acting with function calling can easily exhibit a range of behaviors equating to agency. For example a self diolauge chain of thought prompt technique can get the ai talking to itself, and with layers of function calling, memories like MEMgpt or modern gpt 4 memories, and RAG knowledge graphs it can effectively use real information to make real decisions in an unsupervised dynamic chain. What about that *isnt* agency in the real world? What really matters in the output properties is the prompt, and RAG has a prompt method behind it, so does self discussion chain of thought, etc. The prompt can plan behaviors that arrive at independent decisions comprised of too many system elements too dynamic to generate the exact same output twice. It's more like our brains already than most realize. 😊

Sometimes I enjoy the sarcasm more than the marketing. An LLM is an LLM so it isn't Agentic but the sales guys had an idea. Ship it with an API and call it magic. We like API's. Self-Coding/Tuning Models would be very exciting but of course that wouldn't be an LLM. If you have an infinite number of monkeys that can type at the speed of light (H100s) can you create something smarter than a talking parrot (Not static model but active system?)

future is everyday lately.