ChatGPT Atlas

What value? I haven't used them myself, but from reviews I've seen on Youtube they appear to be flaky and not all that useful. It reminds me of when voice assistants like Siri came out, and it turned out that the only thing they were good for was setting timers, controlling music playback, and gimmicky stuff like that.

Try it yourself and ignore the hype. For certain things it can be really useful, just keep in mind that it's early days.

I recently used Comet to find out of print movies that were never released on DVD/Bluray, then find them on ebay, then find the best value, then provide me with a list to order. It felt like magic watching it work, and saved me many hours of either doing it myself or scripting it.

I did have to repeatedly break it into ever smaller tasks to get everything to fit within the context windows, but still... it might have been janky but it was janky magic.

At what point does the gimmick become a product feature that people use and come to expect? I set alarms/timers and control music with Siri every day. Siri still sucks for more than that, and I wish it were good for more, but I really do like and use those features.

Surely this is something you can do with simple searches...? Unless you're reaching the level of trying to buy dozens or hundreds of movies.

https://xkcd.com/1205/

> Significant losses: Internal documents revealed that Amazon's devices division lost over $25 billion between 2017 and 2021. A separate report estimated the Alexa division alone lost around $10 billion in 2022.

But why? The AI can go off and search and I can do other stuff while it does.

The point is the multiply how much you can get done, simple searches still require me to be present and to do the work of compiling the list myself, this type of busy work seems much better suited to tools like this that take a sentence or 2 to kick off

This xkcd comic doesn't apply anymore due to AI making generating automation code trivial.

I'm mostly just curious how that experience differs from using chatgpt directly and having it run searches and present the results?

My brain can't work like that. When I'm pursuing things, I always await until the blocking operation is done, something about uninterrupted train of thought and avoiding context switching.

Think about this way: They subsidized the hardware and hardware development too much and created a messy third-party ecosystem rather than focusing these 25 bn USD on developing an AI chat (OpenAI spent less in total).

With LLMs we're all becoming managers. Good news is we'll get more done. Bad news is that we'll have to get way better at persisting mid-state process status (I sometimes ask my LLM "could you summarize what we were talking about and why"), tracking outstanding tasks (linear for our agents) and jumping between contexts.

I am also finding work is becoming more tiring. As I'm able to delegate all the rote stuff I feel like decision fatigue is hitting harder/faster as all I spend my time doing is making the harder judgement decisions that the LLMs don't do well enough yet.

Particularly tough in generalist roles where you're doing a little bit of a wide range of things. In a week I might need to research AI tools and leadership principles, come up with facilitation exercises, envision sponsorship models, create decks, write copy, build and filter ICP lists, automate outreach, create articles, do taxes, find speakers, select a vendor for incorporation, find a tool for creating and maintaining logos, fonts and design systems and think deeply about how CTOs should engage with AI strategically. I'm usually burned pretty hard by Friday night :(

This message is so alien to me... I fail to see the real world value of any task you described. Aren't we all fooling ourselves at this point?

everything AI is doable manually. the point is AI saves us mental toil (maybe).

Alexa came out in 2013, long before transformers (2017) or chatgpt (2022). The state of the art back then was closer to CleverBot than anything useful.

We will likely have decent standalone voice assistants at some point soon but Alexa and Siri were way too early for that.

If the movies were never released on DVD/Bluray, what is it that you are searching for on ebay? VHS?

You're discounting the mental effort that goes into describing what you want to an LLM and then verifying the output, which is only worth it if it actually saves you time. Doing a simple search for a few DVDs does not seem worth the effort to me.

You’d be surprised.

I consider myself extremely competent at getting niche results. Couldn’t for the life of me find a certain after market part for a home appliance.

I go ask Claude to find it and it comes back with exactly what I need. One of its queries hit a website with a poorly labeled product that it was able to figure out was exactly what I needed. That product was nested so deeply in results that I would have never found it on my own.

VHS, 16mm film, etc. I enjoy owning films that are basically unobtanium.

> hundreds of movies.

That was the goal, yes. In the end I only actually found about 10 I didn't already own, but the AI had to wade through a few hundred to find them.

It wasn't a few. There are literally thousands of films that never made the leap from VHS to DVD. My starting list was only a few hundred, but still it saved me hours.

I think this is the natural endpoint - local models doing something like what Atlas and Codex are doing, acting like a firewall between the user and web. You don't need to wade through the crap online yourself, the AI extracts the useful signal for you, acts like a memory layer for your values and preferences. Don't like the feed ranking - use your agent to extract, filter and rerank by your criteria. Not a big fan of dark UI patterns? Not a problem anymore, the UI can be regenerated. Need a stronger model? Sure, your agent can delegate.

I see this as a big unbundling, since your agent has your ear now, not Google, not social networks, they lose their entry point status and don't control by ranking, filtering and UI what I see or what I can do. They can spread out searches to specialized engines, replace Google for search, walk above all social networks and centralize your activities so you don't have to follow each one individually. A wrapper or cocoon for the user, taking the ad-block and anti-virus role, protecting your privacy and carefully reducing your exposure to information leaks.

All of this only works if you can host your model. But this is where the trend is going, we can already see decent small models, maybe before 2030 we will be running powerful local models on efficient local chips.

We are on exactly the same page. I wonder if anyone is working on something like this? It would be incredibly meaningful work, I think.

This is actually along the lines of what I'm working on in my free time at the moment. I am working to extend a local model's memory to allow smaller self-hosted models become a better solution than paying someone else.

Once this is working better, it will allow to extend the abilities of local models without running into the massive issues with context limitations I personally was hitting for self hosted.

> the only thing they were good for was setting timers, controlling music playback, and gimmicky stuff like that.

That's _still_ the only thing they're good at...

> acting like a firewall between the user and web.

A firewall which can set the user on fire.

https://guard.io/labs/scamlexity-we-put-agentic-ai-browsers-...

I'm convinced it's some kind of SV-induced "culture of hyper-productivity", the only apparent goal is to be _more_ productive. Endlessly. There's no goal, or end-state, or reason, other than to just be productive. Downtime? Wasted productivity time! Hobbies? Wasted time unless you can monetise them! Get an AI to do it, and be _even more_ productive!

Similar experience. I searched all over for the install manual of my HVAC system (~25 years old at this point). Multiple days turned up nothing. Chatgpt gave me a link to a pdf that Im pretty sure was supposed to be private in about 30 seconds.

I am actually. Let's further discuss: send me an email at this temporary address cdvcrg8zu@mozmail.com to keep in touch.

I shot you an email from my vanity domain. Feel free to edit/delete this message.

I hate to tell you, but Siri and other voice assistance are not good at controlling music playback. I have been waiting for that specific feature to get “good” for over a decade now.

The closest one came to handling controlling music playback well in anyway was Cortana but even Cortana didn’t do the things that I needed to do with my voice while controlling music play.

The biggest Used case for me was always hey Cortana hey Siri add a specific song to now playing and play it next. No matter what on any operating system. The voice assistant delete the entire queue and then probably plays the wrong song. Not only will they play the wrong song, but they will play the entire album from the wrong song if it is available so now I’ve gone from a playlist that I was building and listening to on loop for potentially days to some album that I don’t wanna listen to because it could not just add a song next.