This is best Albert Gu interview. Full stop!

This speaker isn't providing a definitive opinion on anything. For every question, her answer sort of start like there isn't any wrong answer.

he looks so like richard hendricks🤣

excellent interview. great questions.

The current epidemic of misinformation is an existential problem. How can we build an open source and decentralized global platform for fact checking that can explain how we know what we know?

It will be interesting six months from now to see how the platform for Capital One becomes part of an emerging global platform for collective terrestrial intelligence.

Excellent discussion, enjoyed this.

Very grateful for your work in this field. A question; in cases of Gluten Sensitivity, one's own immune system, as I understand it, targets one's own body/cells, caused by the presence of gluten in the food one has ingested. This, one's own immune system attacking one's own body/cells seems to upend the very reason for the immune system (?). I realize that this field is a highly complex one, having numerous branches of studies, but I hope some progress is being made in this fairly recent diagnosis in the medical field.

great insights, like this a lot

this dude knows what he's talking about, completely agree

I couldn't have said it better myself 🔥🔥🔥

Great discussion! Very insightful.

He gives a hands-on review of his confrontation with tacit knowledge vs explicit policies that every analyst runs into when trying to encode business policies into an automated system at 00:38:57.

Could we get to the same point by creating knowledge Graph from our knowledge base. then extract triplets. Then embed the triplets in multi-demensional space

Wonder if this is the same guy who coded Zoo Race.

I completely agree! Especially with how important machine learning has become to the development stage of forming a hypothesis for much of the life sciences.

(16:40) "Benchmarketing" 🤣 So true!

Amanda Askell is the crush we all never knew we had until her thesis ❤😍❤😍❤😍❤❤

I have a bit of an ambitious request, but it would literally be game-changing, for me and many others, if you could somehow make it happen. I have tried reading the JEPA paper (not the V-JEPA one, but the general one, the "A Path to Autonomous Machine Intelligence" one), but a lot of concepts are way above my level. Could you, somehow, get the man himself-Yann LeCun-to discuss the main ideas in this paper? He has done a lot of talks on JEPA in general, but often he dives deep into the details without making sure the core concepts are clear (and I think this is the main reason why he is so misunderstood in the AI community). With enough time (and the right interviewer 😉), I think it could help clear up a lot of those blurry spots. For instance: 1. What is the difference between "reasoning" and "planning" (according to him)? 2. What is the difference between just "planning" and "hierarchical planning", if any? 3. What is "persistent memory"? Isn't that something LLMs already possess (since they seem to "remember" things from their training data)? 4. In the JEPA paper, he talks about "short-term memory". Is that different from "persistent memory"? 5. The idea of JEPA supposedly "understanding the world" by focusing on the bigger picture and ignoring details makes sense to me. But is that reasoning, planning, or something else? 6. He recently started talking about DINO. Is that related to JEPA? Can DINO plan? These are some of the questions I still struggle with, even after listening to his talks over and over. I honestly don’t think it’s even worth bringing up concepts like "regularized methods", "representation collapse", "contrastive methods" and other ridiculously abstract ideas until the core concepts are made clear. I also love when he uses analogies or real-life examples to explain his points, so if you could have him include some after each explanation, it would be even better! Speaking about "real life", he often points out how animals can do a lot of things that LLMs cant. To me it's obvious he isn't talking about stuff like "ability to do math" or "speak language" but it would be great if he could explain his thoughts in more details. So my question on this subject would be: can animals reason, plan, or do both? Could he provide concrete real-life examples of animals reasoning and planning? Of course, I am not delusional. I know asking Yann for an interview and specifically covering some of the topics I mentioned is a pretty big ask, so I will 100% understand if it’s not feasible. This interview is already very good. Take care! (btw, this is a repost. Sorry for the light spam)

The WiFi radios in mobile computers would be more fully utilized if the devices could create ad hoc, mesh wireless networks, materializing local supercomputer clusters. Smartphones with Qualcomm SoCs might comprise the largest number of such devices. The processing could be utilized in generating embeddings and in extracting entities and relationships for constructing and querying shared hybrid vector and graph databases. Even more processing is needed for entity resolution and deduplication, fact checking, and for calculating scores for trustworthiness of stored statements. Building an open source and decentralized global platform for trustworthy collective terrestrial intelligence might be one of the highest priorities for humanity.

im trying to get ChatGPT to work with Claude Engineer

Asking the model to do too much is a sign that you're not thinking it through enough yourself. Good observation.

This woman is awesome!! She lowkey just shared the formula to disrupting billions of dollars worth of data business 🙌🏾💜

Great Podcast, how can we contact her to ask some questions.

very nice already saved

Hi, may I know how to add your channel to Apple Podcast?

In long conversations and conversations that resume after a period of time, the solution is integrating the voice agent with long term, structured storage, probably a hybrid of vector embeddings and entity-relationship graph representations. This is crucial for collective terrestrial intelligence.

There have been reports in the news the last couple of days about Whisper inventing comments while transcribing speech in health care settings.

Just listened to this incredibly helpful discussion about agents. Thank you @samcharrington for always asking for examples - it adds so much to the conversation. Chase Harrison gave such a great explanation of the Langchain universe and how they seek to solve the frustrating problems of building with LLMs.

Very instructive and inspiring talk

V & V as the start of safety critical? Are we "testing quality in"? What about review of the design itself? Where did hazzard analysis / fault tree go? Where did FMECA go? Increasing complexity with an opaque, training-data based approach is gonna explode the risks in the other analysis.

❤️🍓

No alphabetical exodus married to indo European language to build English had to by pass homer out of context. This perfect alignments was needed to better explain realty to put it back right. Took double use hi/ bye, yes no to sloppy loose generalization of feeling word couldn't reform or make a case for freeing ourselves from servitude. Their is no evolution except in metaphysical memes.but even its the math mapping formalism blueprints in the book of jobe. Thermodynamical time. While babylonian greek assyrian me-assure is holographic big bang model evolutionary mythology cosmogony we uses to model. Darwin repackaged spiritual essence from plato borrowed anylitical phylospher blake multi verse arguements. The loss og Hebrew / aramac accent greek letters adaptation. So would the universes full knowledge or data hitting us in the face right now not leave us to learn all intellectual messy complex fluidlike dynamics? Meaning that we began with full and are looking to dig out hidden axioms of complexity put it in our world tech and material sciences then society adjusts or we attempt to form and shape around it ? In my mind super intellect is all that can be archived and this doesnt allow for anything other than jibbresh in visual spectrum. But hooked up directly to all spectrum would have no reason to reason at all. It could have logic like dna in that we can turn it on off make it mutate or mimic. detail all 5 sense to take advantage of miss aligned measure deformity and or hulucinations to benefit needs and demands as super positions. Increasing vision in cameras resolution. Cant clean black body Noise out of systems its limited restraint per richard finneman if i remember it correctly . Artificial intellectual property legal terms appear accurate in English . But it does expose our flipped notions in evolutionary 4th dimension time umbrella term stochastic approach. Itt mind math mapping formalism also reads back how our tools of approximations only model interactions. Which then says morphological thermodynamical is knowledge, made of intellectual fluidlike bits parts and electrodynamical intelligence about those. It was masonic lodge for men eastern star for woman archived data. We call them separatist, puritans like William tyndales burned alive in Brussels. Pilgrim because they help invent archeology to confirm it. Jews lost their language, no offense to Europe it did get letter for letter translation good better than the eastern indo European they was completely lost in translation and out of context in comparison to early alphabetical exodus married to indo European language encryption plagerized correlated prescribed back upon the world around them

Oldest AI podcast I know!!

Here are the key takeaways from Simon Willison's interview on the TWIML AI Podcast courtesy of NotebookLM: ● Willison primarily uses LLMs for coding, leveraging their speed and ability to generate code in various languages. He highlights that hallucinations in code are less problematic as they are immediately detectable upon execution. ● He emphasizes the importance of QA when using LLM-generated code, as they often produce errors. Willison stresses the need to test, rerun, and thoroughly examine the code for potential issues. ● Willison finds voice interaction with LLMs to be incredibly productive, allowing him to code even while walking his dog. He uses ChatGPT's voice mode and code interpreter for this purpose. ● He demonstrates the power of LLMs with tools by detailing how he created a C extension for SQLite using ChatGPT's code interpreter, highlighting the model's ability to access and utilize system utilities like GCC. This example demonstrates that LLMs are capable of tasks far beyond basic code generation. ● Willison sees LLMs as powerful prototyping tools, enabling him to quickly test ideas and create UI mockups using Claude's artifacts feature. This rapid prototyping significantly speeds up his development process. ● He advocates for constantly exploring the capabilities and limitations of LLMs by trying them for various tasks, even those seemingly outside their purview. This experimentation helps understand what they excel at and where they fall short. ● Willison acknowledges that LLMs can get stuck in repetitive loops when debugging, recommending clearing the chat history and starting fresh in such situations. This highlights the importance of managing conversation context when interacting with LLMs. ● He emphasizes the effectiveness of providing LLMs with code examples from previous projects to guide them toward desired outcomes. This approach mirrors how he would guide a human intern, providing clear direction and context. ● Willison notes that while vision models are promising, prompting strategies for them are still underdeveloped compared to text models. He believes further research and experimentation in this area are needed. ● He expresses excitement about the potential of Open Source models like Llama but finds the hosted models currently offer superior quality and affordability. He acknowledges the rapid advancements in Open Source models, particularly Llama, and expects increased competition to drive down costs. ● Willison believes focusing on current capabilities is crucial while acknowledging the potential future advancements in LLMs. He anticipates significant releases from major AI labs and hopes for progress in allowing selective disabling of safety filters for specific use cases like journalism.

The research on people reaching more accurate mental models of reality by observing digital agents debate with each other is VERY interesting.

Perhaps a platform for collective human and digital intelligence would be more immediately beneficial, and less risky than AGI. The platform would merge knowledge and sentiment expressed in conversations with people around the world into a shared digital representation. It would fact check claims. It would compare action proposals with the principles of the Universal Declaration of Human Rights and the Earth Charter. A hybrid platform of language models and vector and graph databases such as in the LightRAG project might be a step towards collective terrestrial intelligence.

Hello sir, First I am sorry to say that without your permission I messaged you. I am a Digital marketer and podcast promoter. I know how to increase followers, downloads, and traffic for any podcast. I will keep your Apple podcast top ranking chart (1-20) in your country. Would you please give me a chance to say more details to you?

Robert Downey Jr. is here 😮

o1

Super happy to see you on YT! Been missing you since Alphabet scraped Google Podcasts! Awesome content.

I learn so much every time I hear Shreya speak, thanks for having her on!!

❤️

Nightshade's effects remind me of a story where a guy ruined an AI image generator by asking it to generate Kermit so much that Kermit started appearing on every image that the AI created.

The code run or not already is the verifier. Just train model with code thar can run. Ou use a simple agent to check if the code can run. That's why LLM are greater to write common code. In other hand, there is no framework to check if some sentence is true given the specific context that it applies. So, while this is not invented, the LLM will appropriately sent gramar correct sentences but 10% of inventions.

Make sense life is complicated

10 minutes in, great conversation... Instant subscribe.

Wilson Kimberly White Jennifer Moore Michael