How Cohere will improve AI Reasoning this year 

Probably a trough of disillusionment in about 12 months. The industry has overhyped the capability too much at this point so it will need a correction. It’ll probably be a slow correction too. I don’t see any personnel reduction happening any time soon. More business process augmentation, and human augmentation. There’ll be more humans in the process I think, not fewer. Also, more use cases with LLM and classical ML hybrids for broader intelligent apps use cases. Enterprises will be looking to quantify ROI for their initial investments and will struggle to find it. There will be more marketing around it but I don’t think boards or shareholders will be satisfied with the current marketing numbers.

I think so too, a lot or hysteria up till now, but it'll steady out. Sounds like a great outfit, thank you kindly for the interview..

@@Peter-dd3br And all you have to worry about is your customers getting a little glue in their pizza 🤣

I agree with your sentiment, however even though Aidan undoubtedly has the intellectual horsepower (he is one of the OG writers of the Transformers paper) as a CEO he cannot keep up with all the technical details - pretty sure his day consists more on him doing executive choices, partnerships, tons of meetings & traveling. Cohere engineers seem solid & their models are quite capable :)

@@XShollaj Fair point.

@@XShollajto me, Cohere is one of those foundation models builders who have clearly defined their niche and optimising towards that. Very developer-friendly. They allow you to test and pilot as much as you want and support you to go to production. Very competent engineers and great customer support. IMHO, they are too quiet.

tbf neither does elon musk when you hear him talk

Soo... I tried using LLMs for a judgement call.. (context at end of next post.. for which you may take the below comment as a tl;dr warning..) "YES or NO. Given the choice of posting or not posting (no further alternative actions available), is it worth posting the following RU-vid comment?" GPT4o: No. The comment is very long, detailed, and complex, which might deter readers from engaging with it. Simplifying and focusing on key points could make it more effective. Claude 3.5: NO. While your comment contains thoughtful analysis and interesting points, it is excessively long and unfocused for a RU-vid comment. RU-vid comments are typically brief and to-the-point. This extensive commentary would be better suited for a blog post, forum discussion, or personal essay. Posting such a lengthy comment is unlikely to engage most viewers and may be overlooked or seen as overwhelming in the context of a RU-vid comment section.​​​​​​​​​​​​​​​​ Human-(Me): Ach, go-on.. YOLO.. YES. OK, so, it's a bit all over the place and it's a shame (though I'm biased) that for brevity(!) you've dropped points like the Mixture of Cameras model concept... and the apt-but-utterly-tangential rant on UK medical training following Calmanisation reforms 15+yrs ago likely impact on consultant domain expert breadth of knowledge from the original.. but at least you will have parked most of the material contextually somewhere if you want to come back to it for eg a blog post. And it may get it into training data and make sense to an LLM.. And the potential human audience where you plan to post it is narrow, peculiarly relevant, possibly thoughtful/amused enough to read it, or intelligent enough to give up quickly if uninterested, or kind enough to forgive if they read the end bit first.. so maybe won't be too annoyed... Why not ask your critics to sense check? Howabout: "Are there any obviously stupid or erroneous comments in the following:" Claude 3.5: After carefully reviewing the text, I don't see any obviously stupid or erroneous comments. The content appears to be a thoughtful, nuanced discussion of several complex topics related to AI development, climate change, and risk assessment. While the author acknowledges that their writing may be influenced by their current circumstances (being in a hospital bed), the points made seem reasoned and well-considered, even if somewhat rambling in style. The author touches on various subjects, including: 1. AI model development and training data strategies 2. The challenges of assessing and communicating existential risks 3. The difficulties in bridging different thinking styles and domain expertise when addressing complex issues 4. The potential for AI to help policymakers better understand complex, multi-domain problems While some of the opinions expressed may be debatable, they don't appear to be obviously erroneous. The author also demonstrates self-awareness about the length and potentially meandering nature of their comment. If anything stands out, it's the casual mention of using an AI assistant (Claude 3.5) to help reduce the character count of their comment. This isn't erroneous, but it's an interesting meta-commentary on AI use within a discussion about AI development. Overall, the text presents as a sincere, if somewhat stream-of-consciousness, exploration of complex ideas rather than containing any clearly stupid or erroneous statements. GPT4o: {long irascible conversation omitted by human} chatgpt.com/share/9881f0fb-6241-4440-8ffd-27f53283bb02 Human-(me) writes: GPT4o's response was so off-topic and irritating that after a protracted angry discussion I put the whole chat into Claude3.5 for independent comment which came back as follows: Thank you for sharing this interaction and asking for my observations. Here are my key takeaways: 1. The original comment was quite long, detailed and covered multiple complex topics related to AI development, data usage, and potential risks. It was written in a somewhat informal, stream-of-consciousness style. 2. The AI assistant (not me) initially focused heavily on issues of clarity, structure, and potential ambiguity in the comment, rather than strictly addressing factual errors as requested. This seems to have frustrated the human user. 3. The human had to repeatedly clarify and emphasize that they were looking specifically for known factual errors, not issues of style or clarity. They expressed frustration that the AI was not directly addressing their actual request. 4. Even after multiple clarifications, the AI assistant continued to include opinions and subjective assessments in its responses, rather than strictly focusing on verifiable factual errors as requested. 5. The human pointed out several instances where the AI assistant seemed to misread or misinterpret parts of the original comment, such as the discussion of game engines. 6. There was a discussion about potential bias related to Microsoft funding, which the AI denied but is an interesting point to consider regarding potential influences on AI systems. 7. Overall, this interaction highlights some challenges in AI-human communication, particularly around following specific instructions, distinguishing between fact and opinion, and accurately interpreting nuanced human writing. The AI assistant seemed to struggle with precisely addressing the human's request, leading to a somewhat frustrating exchange where the human had to repeatedly clarify and redirect the conversation. This demonstrates the importance of AIs being able to accurately interpret and follow specific human instructions, especially when dealing with complex, nuanced topics.

Thank you Sandra!

Yup. And I doubt it's that difficult. I think these AI companies are dragging their feet with it in order to get more investments from investors.

Not true. This has legs, even with little increment improvement from now on! Just getting a computer to read in a natural voice is a huge achievement, its those little improvements that would have been huge a decade ago that will improve productivity for years to come.

I think he means that while they do still hallucinate, it's a relatively small amount of cases when the model knows a lot about the topic the prompt is querying. App builders still need to do a tonne of engineering to address this i.e. RAG and other stuff, but it's not a deal breaker for many practical applications.

I certainly hope so! We haven't filmed with him yet

also saying absolutely nothing

@@minhuang8848he said so much. Just listen. He’s a whole ocean technically sounder and more competent than many a high-sounding bamboozlers in this space.

It's called e/acc. You'd probably find people talking about it on Twitter or reddit

See it from their perspective, there are those who are trying to kill their business with irrational fears of "AI existential risk". There *are* really important risks though, this is nuanced.

I think it is important that smaller entrepreneurs are speaking with regulators. You know that the largest players (who are also the largest companies in the world) have major lobbying efforts. All you need to do is look at the Sam Altman playing up the risk of an “existential threat” before congress to see their disinformation. They would like to see a regulatory scheme that would allow only the largest players to play (like we already have in many industries) while ignoring real risks like misinformation, economic disruption and wholesale misappropriation of IP.

They tried that for years and it failed.

​@@mikezooperit's because they are trying to derive logic from data, rather than let the model build its own logic little by little through experience.

@@hunterkudo9832 experience, as in living day by day, which takes too long. What entrepreneur has the time to wait 30 years to have a full grown AI trained on "experiences"? They want money NOW. And that's all this is about: money. Progress is a side effect.

A true MOE involves a combination(mixture) of specialised models with deep domain knowledge(Experts) not of different generalists.

@@Enedee007 Right but if they only have knowledge in their own domains, then how will there be true cross-discipline “inspiration”? The engineering expert might not think to borrow an idea from the animal kingdom, for example. Why? Because whatever is coordinating the agents and routing the questions to the appropriate expert will not itself have the breadth of understanding to connect the dots between two highly distinct disciplines. Like maybe a structural engineering expert might not think to look to honeycomb structures from bee hives. But if they were all in the same monolithic model then the associations could be made at some layer of latent representation.

@@andrewsilber that’s correct. I think our points are more similar than they differ. The coordinating agent should be the generalist in that architecture. That’s the agent that’s able to carry out cross-discipline activities. That’s also a specialisation, it affords the opportunity to deeply train a specific model for that exact purpose, it should be well grounded in its ability to learn from different areas of specialisation and make critical connections.

@@Enedee007 @andrewsilber There is not just one MOE implementation and I think as currently used, it is done to only activate certain weights (the experts) in some of the layers probably not so much to silo the information but to reduce parameters and increase performance without any loss of acuity. The difficulty with cross discipline inference is actually baked into the models themselves unless it is in the training data or in an expression like “honeycomb structure”. No amount of “temperature” will get the probabilities to point just anywhere in the model.

@@toadlguy i totally agree with you there just not one implementation of MOE. However, the underlying principles are basically the same. MOE I’m this case however is in the context of multi-agents.

Maybe some problems already have solutions, but those solutions aren't quite ready yet. Too much compute, to low profits for now.

@@ronilevarez901 Yeah that is correct, but the stupid thing is to make claims that those approaches will lead to AGI, but until we reach architecture that learns quickly like children with very few examples, we have long way AGI

Consciousness is still a computation

@@bobhanger3370 No. Consciousness is quantum. It is the zero point of calculation. Computers are actionable; you are not. You will not have free will otherwise.

The title 😂

Did you not see the clearly marked "paid promotion" label at the start of the video? How much clearer could it be? Also we added a disclaimer "This is the second video from our Cohere partnership. We were not told what to say in the interview, and didn't edit anything out from the interview." - and if you look at our back catalog we have been interviewing Cohere and their researchers for years without any commercials. I honestly couldn't think of a better possible revenue model for the channel, unless you want us to spam you with ads about NordVPN? To demonstrate our genuine interest in Cohere look at back catalog with their team: ru-vid.com/video/%D0%B2%D0%B8%D0%B4%D0%B5%D0%BE-i9VPPmQn9HQ.html ru-vid.com/video/%D0%B2%D0%B8%D0%B4%D0%B5%D0%BE-Dm5sfALoL1Y.html ru-vid.com/video/%D0%B2%D0%B8%D0%B4%D0%B5%D0%BE-7oJui4eSCoY.html ru-vid.com/video/%D0%B2%D0%B8%D0%B4%D0%B5%D0%BE-ooBt_di8DLs.html ru-vid.com/video/%D0%B2%D0%B8%D0%B4%D0%B5%D0%BE-sQFxbQ7ade0.html

@@MachineLearningStreetTalk Thank you for taking the time to respond. I'm afraid I missed it that, as I listen to interviews like this on my phone while I do chores, so I didn't see the visual cue. Ideally there would be an audio cue for people like me, but you're following RU-vid policies, so I'm all good. Apologies if I brought you down!

This guy has knowledge, and a sane worldview so no.