Відео

Thank you so much. With your inspiration and the help of Cline I was able to create my own MCP-Server that reads/writes on MQTT topics. With this I'm able to tell Claude settings of my home automation.

does gemini vision has an api? or do you have to browse the website you want to scrape with gemini.

Thanks for the breakdown! Could you help me with something unrelated: I have a SafePal wallet with USDT, and I have the seed phrase. (alarm fetch churn bridge exercise tape speak race clerk couch crater letter). How should I go about transferring them to Binance?

Mess is real! thx

Thanks for sharing! I learn a lot from you 😃 One request that I have is if you could use a better microphone as when there is noise I cannot hear you so well 🙏

Good one, I actually think it is good to see how you debug and resolve issues. Keep it coming please, thanks from Ethiopia.

Thank you for sharing!

אח יקר פרגנתי בהרשמה ופעמון :)

Next I have to figure out how to use Claude desktop app on Linux.

Great Video! what's the name of the tool that converted your speech to text for Cline?

What voice to text tool do you use?

of course the debugging... of course the mention of me here now would have you then say you may make a debugging video to counter what i was going to say in that of course the one type of video i want to see, debugging and actually doing things, ... anyways, fucking time travel psychosis insane impossible reality wehatever hail .... janything but me i guess whatever. Oh and this is cool. -_-

Hopefully they will add the ability to have both audio and text in the same session - It's great to be able to have a conversation, it is very natural, but when you ask it to tell you the code you get "backquote backquote less than html greater than backquote backquote" etc. It would be a lot better if (when) it knows to put output into a text format

Thanks Great video without The fluff as usual. What was the voice typing tool you were using?

Hey Yaron - thanks for always being on the ball testing the latest features! Two quick questions: 1. You mentioned openai has it's own equivalent of MCP - what is that? 2. can other non-claude models use MCP? E.g. Llama3.3 or even gpt4o? Thank you!

Thank you brother for this incredible news! OMG I just used it and it is so crazy powerful that I want literally use O1 pro + MCP/Cline in chain...... I cant find anything else on it yet.

Good start, but seems very limited in comparison to what a tool can do. It can only scrape what it 'sees', so much of the info on the page is cut off. The Airbnb test you did is an example. With how powerful Gemini is now, you'd probably be better of doing a Ctrl-A on the page, dumping that into gemini and having it turns the info into JSON. Good video though, to show some capabilities.

It's kinda lame it always stops after a couple minutes of operation. Its kinda useless for serious usage. Have you gotten around these limits? Notice your demos are all very short. Did you plan on mentioning this. You scraped air bnb? You got one line of one page. This is hardly efficient even if the demo they have now didnt cut off after two minutes.

Volume very low

that's creative as fuck

Great content but a horrible microphone.

Awesome!

Mindblowing for real. Wow.

thanks - new gemini looks great:)

All AI models are rapidly being updated! It really feels like AGI is coming soon!

I've created a follow up video, showing how to use the new gemini model for web scraping, check it out here: ua-cam.com/video/8oD7juOSEN0/v-deo.htmlsi=W0d0CD3tAxg5WTax

wow that's insane

How can you access this?

Tutorial on building with it from GitHub

I am LOOOOOSING MY MARBLES ON EVERYTHING

Thanks good video I'm glad you left the restart in there as well. Just reading forces why I don't understand how the start-ups like cursor Bolt have been able to secure funding. All of these developments by OpenAI Google anthropic etc are going to steamroll them. Having said that I still find I use cline the most out of any of my tools

Yep, your sharing some good stuff. Ive been messing with the 1206 LLM for coding and its pretty good but this is pretty amazing. I think this real-time video tech could be good at detecting people's emotions and maybe catch them before they "may" do something wrong or dangerous, but that could also be horrendous as well lol.

great overview! thanks

Of all the AI channels, yours is the most cutting edge. Congrats!

Thanks for exploring the ReasoningAgent, we loved the comparison with o1 model, and we are iterating on the approaches to make this better and better.

It's the QA Fishbone process, usually used for root cause analysis - complete with brainstorming results at every step. Cool. Thanks.

You're right that this is a very important problem. If we had a really good grader then we wouldn't need the tree, and their prompt seems way too simplistic. I've experienced this "always high" scoring before and I think the problem stems from the fact that the LLM wouldn't have made the suggestion if it thought it was bad. I don't know if it would change their logic completely, but it might be better to let it pick the top X suggestions, then explore those, pick the top X, etc. You could also come up with several criteria for it to use to judge by and average those, but that might be difficult to implement in a general purpose framework because those criteria would likely change based on the problem.

Very interesting. I wonder how a relative scale (rate brainstorming candidates relative to each other) would perform or a ranking mechanism would perform?

Here's a follow up video I did about the reasoning agent. Exploring how to make the most out of it by optimizing the agent grade system prompt: ua-cam.com/video/AhQtNAAccxI/v-deo.html

Very interesting! I think I'm going to try it.

How would you create the best LLM teacher or tutor? Does any LLM exist right now that is specialized on teaching or tutoring? A tutor that is trained on pretty much all of the existing pedagogical tools, processes and frameworks and can combine them together to create the ultimate teacher.

this is slow af, and it doesnt always work and not to mention expensive

You can also create these modularly meaning 4 or 5 hooks, the same middle and different calls to action....thinking of it as a matrix - say 4 x 1 x 4 = 16 parts or different ads

Fire!

Thank you for this video! I've gotten a lot of value here ❤🌟

I admire your perseverance. Anyone who codes has been exactly where you were. Post the fixes, video and code! :)

Great job! Two comments. 1. The package actually contains a green powder that mixes with water and turns the mixture green instead of being a clear liquid that is consumed directly. But, of course, the model did not have this information in context. So I wonder if you could solve that by including that information in the prompt somehow? Maybe by another video, or another picture or somehow using text. 2. Is there any way to use agents to automate the video creation process itself? Full pipeline.

At 13:39 i noticed you are using gpt-4. I wouldn't recommend using that one, its a legacy model and you could be paying more than gpt-4o. Also gpt-4o-mini is most of the time more than capable enough for most tasks. (I am assuming it noted down gpt-4 because you generated most of the code using AI and they have a memory cut off date) I do want to add that I used to be skeptical about AutoGen but you convinced me to try it out and see what it brings! Thank you for that and keep up making videos

I really liked what you did at 28:41 by enhancing the prompt. I usually tell o1-preview to use a list of prompt engineering tactics to formulate the system prompt. This in some cases enhances the output by 200 to 300 percent.