Видео

Yes, both mention problems now in progress. One models woth self-reflection, and second the agent framworks where actors(dev, reviwer, debuger, QA and devops) all have separate context

All these tools currently produced by various companies are just novelties and nothing more for people who want to start their programming journey. In my opinion, it's better to understand the basics first to know what these tools actually produce. There are plenty of opinions online where people go on about what they've achieved with tools like Cursor AI or Replit AI. However, from my perspective, Replit AI doesn't stand out in any way and, in fact, is even worse than Cursor AI with Composer. However, in my experience, when dealing with a slightly larger repository-not a huge one, but one with around 150 files with already written code-the attempt to add a new feature that would use the existing logic and previously prepared tools is more frustrating than helpful. Cursor AI supposedly indexes the repository, so why do I have to point out the exact files it should use every time I want to add a new feature to create something with the previously developed tools? Otherwise, it tries to write something itself, and in a much worse way. There's still a long way to go for these models (a much larger context window is needed to really grasp what’s happening in a project) and for the developers who create these tools. Every time, I get caught up in the same thing-posts on X by marketers or people who are being paid for it (because how else to explain that these tools are still so poor), praising how their productivity jumps by 150% when writing in English. When I provide too many files, these tools start to get lost in what they should do. If there's a specific approach to creating services, repositories, DTOs, or utilities in a repository, I would expect such tools to stick to that rather than suggesting their own solutions that deviate significantly from the standards presented in the repository. Once again, I feel disappointed.

I was thinking a similar thing the other day while using cursor. It had got into a doom-spiral and if there was a little 'scissors' icon or something where you could just cut the chat so anything above it was 'forgotten' by the model but you (as the user) could still read back - that would be _really_ helpful. So much nicer than doing a completely fresh chat and losing the context of your own thinking along with the model outputs. If you could also remove the 'snip' (assuming the model got itself unstuck) I would guess it might also help the model see 'oh, we a that problem and then I did *this thing* to fix it.' rather than getting stuck in the same loop again 10 minutes later.

What I'd love to see, and I haven't found anyone who's done it, is a comparison to Claude projects. A lot of the features seem to overlap.

Great video. Thank u man

I love how your example of replit's supposed broad problem solving scope is a web scraper which is the example in every other AI agent demo

For me Replit only works when I use Python without any implementation of LLMs APIs - at the moment for me useless - maybe for some basic productivity apps or data visualization - I just waisted 25$ for subscription.

I doubt that opportunity will be available for senior devs as ai will also be able to answer all those questions. From the next model stage say gpt 5 opus 3.5 lastest the models after that senior devs going to have to focus on how to stay ahead

I’m so curious if Replit was made using Cursor

Love your observation about the market opportunity for senior developers to help newbies. I’ve been using cursor for at least months to program in variety of new languages to me. Especially before sonet it frequently got stuck going in circles and I ended up hiring a couch/teacher on Upwork to help me get unstuck. It wasn’t super easy to find someone and I have the market for helping new programmers develops with new marketplaces, service offerings, and tools.

token limit seems to be a bigger problem running locally where you don't have 16GB+ of VRAM and you don't trust putting your code in the cloud. For example, was trying to rewrite a 300 line SQL query but it's kind of a one shot due to token limit and 4-bit vectorization, the model just couldn't do it properly.

sooo . . . . you got paid for this endorsement? There have been at least a dozen online and offline agents for Claude 3.5 like this, replit is not even the best one around.

DId you use riverside to record this?

nice video are you open to new ideas on your RUclips channel videos ??

do you think that AI will be able to do these things in the next 50 years? or its a thing that no mechine can compete with humans just like emotions

100% agree. In my experience the only time I have gotten solid results with AI is when I knew exactly WHAT I wanted. You can't leave the requirements up to AI because it can implement all possibilities with equal facility, which is to say it has no true preferences and no "insight" as to what you should do. If you tell it "do X" it will laud you on your fine choice. If you immediately turn right around and say do the opposite of X, it will applaud you in your decision.

Indeed. Turns out that programming in natural language is still...programming! Also, reminds me of this comic: i.redd.it/bi526dh4cuca1.png

Nate, I can't watch this, because of the sing-along subtitles. This makes it impossible to read anything else that's on the screen. It's hostile to people with accessibility issues, even as simple as people who don't speak english - do you expect them to watch with two layers of subtitles on top of each other, one of them jumping and screaming for attention? RUclips has proper, accessible subtitles we can enable. I know that research supports it - this "works" as far as attention retention. But it is a cheap way to improve your stats. Do it with great content instead!

Thanks a lot for this lesson! Excited for the next part :) really liked the "blank page problem" insight

AI will also help us to develop better questions, which may require complex reasoning to ask/answer

You mentioned the models are fallable. What are your thoughts on Constitution AI, where AI creates reward system for fine tuning? Do see significant opportunities to manipulation (bias towards corps and not people) to the base examples?

How can devs use it to collaborate on BI projects over their data architecture? Can it be integrated into common business tools for this function? Cursor doesn't feel mature enough for businesses to put it at the core of their way of working and derive value from it. Like ChatGPT, I feel it needs a Microsoft to buy it and integrate it with their platform before it changes how we work

Huh, I've been using Cursor since it was first rolled out. I thought it was already well known. Nice to see I was so far ahead of the curve.

I would tell my 25 year ago self to pay more attention to math, and never get lazy about performance of your code

That Trello clone isn't that "simple" of an app. I really impressed they could pull that off so easily.

this AI tech makes me think that we will need less developers but those developeras must be senior level since only complex projects need actual humans and the less complex projects need prompting from people who know what they must accept or reject. few but more experienced developers will survive the hunger games. unfortunately I never had a job in tech even though I'm better than mid level developers. just unlucky.

Expectations will continue to be unrealistically high. Estimates are that approx $1T was spent on AI infra. That should result in ROI that impacts GDP. But that's still further down the road, mostly for non-technical, organizational reasons. Keep up the videos!

On the California bill - as the meme has it, why not both? I more or less agree with the idea that we aren't handling either the terminator possibility (which is remote) or the disinformation, misinformation, fake a person, etc. stuff well at all. Testing, rights to third party pentest, audits, etc. can all be used for both reasons.

I'd venture to say that there are a fair number of programming tasks that I would be willing to wait overnight for ... given that I wouldnt need to go back and correct the work. The AI would be like a sous chef or prep cook getting all of the ingredients ready (utility programs, schemas, unit tests etc) before I come in and assemble it to make the main dish.

What's up with the pointless text and pointless subtitles (that aren't even "sub") plastered all over the video? Just make a normal video silly clown. Channel blocked with BlockTube.

Multi step reasoning may actually be really useful to do complicated work and orchestrate a bunch of interactions with faster models. For example, use the multi-step model to design an entire application and break down the design and implementation into smaller steps, then have it input those steps for you into an AI service like Cursor. It'll be really interesting to see how much better the multi step reasoning models are at design in general. I bet those models could create really fantastic systems of agents to complete entire workflows

Multistep reasoning: software development itself?

I actually love the idea of a smarter-but-slower processor option that can “slow cook” while I’m away or asleep. I wonder what the resource usage would be for that, though. “Good morning! While you were asleep, I figured out EXACTLY what color to paint your kitchen! And it only took 1.8 x 10^60 tokens!”

It's a shame that Strawberry doesn't know how many Rs are in the word Strawberry! :)

I'm interested in seeing how these type of apps evolve when more features are added to the product (hence more complex product). It always stop at the PoC level right now and then you don't hear more about it.

This feels really exciting. Would you have an opinion on how kids who want to "code" should approach programming? I have a 9 year old who expresses interest In coding. I'm assuming learning to code the traditional way is probably still necessary...or am I wrong?

can you please not add these baked-in sing-along subtitles? RUclips has built-in subtitles that we can turn on. baked-in subtitles affect accessibility for non-english speakers, who are forced to watch two layers of subtitles on top of each other. (it's also mildly offensive to those of us who know why this is a trend on so many channels.)

Agree 💯 with your assessment, especially for the Agile teams. We're not there yet!

if you look at the price of shares of crowdstrike it does not make any sense. Crowdstrike stock started to drop off a cliff the 1st of July, and fell dramatically for 3 straight weeks *before* the outage of 19th july. The outage itself barely affected their stock price. In fact, 1 week after the outage the price of crowdstrike went back up by quite a bit. uh?

Crazy to think some companies are still using mild hallucinations/(insert random irrelevant security fear) as the reason why they aren’t even experimenting with basic LLMs. We are now speeding towards an agentic future where some companies will be light years ahead…this space is moving so fast it’s almost hard to believe

This is no different than when Auto-Tune came out for singers. It significantly lowered the bar on how talented creators had to be in order to make music. And just like with artists and AI, the artists that can make great art without AI are complaining about how much AI is enabling lesser-talented people to make impressive works. Now granted the fundamental difference here is Auto-Tune is a technology that doesn't rely, inherently, on being trained on others work. However its development and testing was done by those that, no doubt, are familiar with others works. So it, indirectly, benefits from it via its creators. Generative AIs just move the training a little closer to the code (IMO). But to answer your question, I think AI is a new tool in the toolbox for creators willing to use it. Those that learn to use it well will be successful. And those that refuse, will find they are now competing with new set of people they didn't have to compete with prior to AI. To compare this to a radically different industry, it would be like a really experienced and skilled mechanic that refuses to use modern tools like code scanners, bore scopes, or wireless drills/impact wrenches. Then they complain that those that do use them, get the same jobs diagnosed & done faster. I'm sure some here will argue the "moral" differences, all of which I'm quite aware of. But that won't change the reality that AI is not going away. Training AI on copyrighted material also isn't going to stop. It'll just get trickier and more difficult, if not impossible, to prove. So those that oppose AI instead of learning to embrace it do so at their own peril.

I would take the $230 million number with a grain of salt. Sounds almost too good to be true. But still: even $50 million in savings is significant, so maybe this really is the tip of a really huge iceberg....

Thank god someone is taking a stand against this AI perversion.

This business case comparison between procreate and adobe seems asking the wrong questions about doing the picture making work. AI is not about increasing productivity it's a shift about adding connectivity to the picture. This asks what's the purpose of the picture that has that content? Right now the tech scrapes or copies existing content which does not reflect human cognition well. What's the use case of adding connectivity? What's real about connecting? We have the historical moment of Cezanne's multi-perspective images that opened up abstraction. How does that facilitate connecting structures? Anyone familiar with neural networks knows the importance of that chip architecture whether server farms or mobile devices. But the use case of abstraction appearing to disintegrate 'realism' is not about popping out antirealist tropes. It is a deep question of what connecting data adds to a picture for real use. Pattern making AI is unable to address connectivity properties that abstraction raises.

Interesting that you mentioned workflows. One of the first things I did with early Midjourney years ago was to thrown my creations into Procreate so that I could iterate, draw on top of it, change colours, etc. It makes sense as a workflow. Fast forward 2 years into the future, now using ComfyUI and many models, this still makes sense, albeit there are Loras, embeddings, ControlNets and other tools that can get the AI creation closer to my overall vision. But still Procreate is king when it comes to drawing whatever you want, if you have some skill. My reading between the lines from Procreate's blah blah blah is that they don't want to invest on a saturated marketplace, so they decided to announce that through a PR stunt, also at the same time trying to win more AI skeptics to buy their software... In that sense they might've had some PR success, growth in users, less churn, whatever. It's a phase they're going through and it's not gonna last, as you said. Maybe I'm wrong and Procreate's stance was genuine. It would be noble but still heavily misguided unfortunately. Ultimately, a tool maker can't "ban" the other tool. It's like a hammer manufacturer banning screwdrivers. I've used both on the same woodworking project a few times. I don't care what the manufacturers have to say, I just want to work on my stuff, create things and have fun. I like your insight on how Adobe is trying to go after the casual creators and how they're trying to cater to multiple personas. I might have to change my point of view because as of right now my take it that Procreate doesn't really compete with Adobe tools (I think you're referencing specifically Photoshop)... And Adobe users might slightly overlap with Canva, but just the non-professional designers and "casual creators" that you mentioned. In my head as of right now: Procreate = Best tool for drawing and painting, you need to use on the Ipad. It's used for low-key assets, to have the hand draw feel to it. Photoshop = Editing photos, using the smart fill you can prompt to inpaint and outpaint. Yes, some people use it for professional art, with their Wacom Cintiqs but that's a different niche all together. A professional artist might use PS and a CIntiq as a daily working tool on the gaming or animation industry, but for fun they'll use Procreate at home. Or they'll use Procreate for casual projects and little things (like Google Doodles). As for other Adobe stuff, a professional designer will use Illustrator for vectors, InDesign for layouts, After Effects for motion graphics, Premiere for quick cuts and transitions. No-one uses XD by the way, for apps, web and digital products they'll use Figma. Canva = For small business owners or small companies that don't have designers, or maybe people designing that did not study design professionally or have little to no experience in commercial design. Example: graphic or motion designers can definitely recommend Canva to someone on the HR department to create a poster and put it on the kitchen fridge about the company's catered lunch, but they definitely don't use it. They'll use Adobe tools. But you're right, Adobe is trying to go broad, there's money outside of the commercial application, and AI is becoming yet another slight equalizer. I say "slight", because an artist or designer can make or curate things on a level that non-professionals can't. So it really shines on the right hands... and eyes, and sense of aesthetics, and ability to judge design scientifically. As a parting though: I think Procreate's statement- if it truly came from an ethics stance- is somewhat noble, but misguided because the trends are inevitable. There's a crappyfication happening with all content currently, because companies figured out people can't perceive quality after a certain threshold. In general terms, the "not-too-bad" performs as well as excellent. This is what we see with content creation and consumption on social media for example. Spotify is on the same path: who needs artists creating professional music, understanding about orchestration, music theory, etc... When a programmed bot with Zapier and other API calls can very quickly look at trends and generate whatever crap that WILL be streamed and generate revenue. AI is not gonna replace artists and good taste, it can be used by them to do great things. What AI will do is replace content creation at a low-quality level (but the level that is consumed by the majority of the population and that makes a lot of money currently). Eventually, people's taste might improve and things can get better, but until then, "lo-fi quality" will be the norm.

Nice finally some common sens.

This was very insightful. I'm in HR. You are 100% right it's become the best versions of people so instead of getting 100 resumes and having 5 elite candidates they are all now elite and super qualified on paper. Now, as an HR professional looking for the next role and step up the 🪜, I completely understand the tool and the magic. I think the future will see employers move to a "resume and" economy. Where additional supporting resources,case studies will be part of the process. Very interesting

I've seen some of the popular items try to cite their sources and do so, correctly in the sense that the source has the right information, but oddly the engine did not pick it up and included the wrong figures! Meanwhile, I wonder what the 1A has to say about fake endorsements. They might be protected.

Why do you keep assigning us billion dollar problems, assuming that we are smart enough to solve them? ;-D

Taylor prob won't sue Trump, because his supporters will boycott her concerts and music. I guess, probably.