There are two companies that have lost my trust, probably forever: X and Meta. I don't see myself ever trusting anything coming out of either of these companies ever again
I've been using it for the last month or so. IMO in the same way we went from tab complete -> prompts -> agents, this feels like a next step on that evolution. I highly suspect others will be following suit. I was surprised with how much it felt natural to interact with agents in this way.
Biggest advantage is each one owns its own routines, context, and domain, and they can communicate between each other. Similar to hermes they build out their own skills, but by keeping the bots separated by domains, you end up getting better results out of them.
Additionally though each one has their own computer, which means async work feels like it actually works. I haven't had to juggle worktrees for the last month.
Biggest downsides are token expenditure. I've used more tokens this month than not this month. That's not a typo - I've used less tokens in the last 5 years prior to this month than I have this month. Always on perpetual agents use a LOT of tokens. IMO this is building for the future state where tokens are vastly cheaper, ie in a post-ASIC world.
I wanted to make something that didn't feel like just my logo on a shirt, so I had one of my bots reach out to ~40 fabric suppliers in vietnam, negotiate prices, lock one in, and get samples made. First samples should be finished today. It's been something I've wanted to do for ages, so it was cool seeing it actually happen. The fabric supplier bot worked with one of my prototyper bots to create a randomly generated pattern using my logo, which it then sent as a .ai file to the supplier.
> I wanted to make something that didn't feel like just my logo on a shirt, so I had one of my bots reach out to ~40 fabric suppliers in vietnam, negotiate prices, lock one in, and get samples made.
Isn't this one of the problems foreseen with this? For you, it was a single prompt - for 40 companies, this probably took up some time.
What happens when fifty people fire off a 15-second "get me a shirt" prompt? When five hundred, five thousand, five million do?
Isn't the answer obvious? The 40 companies will have to use AI to filter the messages too.
If your business is selling tokens, it'd be extremely lucrative for you if the whole society relies on tokens to perform basic operations. That's where we're heading to.
A reasonable pushback. It originally only reached out to 5, didnt hear back, so reached out to 5 more. I personally pushed it to reach out to an additional 30 after that.
One of the difficulties of sourcing this is a lot of the suppliers in vietnam are only contactable via whatsapp. Emails are monitored far less. It's one of the reasons I haven't been successful with this in the past despite trying - it's a very word-of-mouth network.
I don't think your reply addresses the point of my comment at all, which is that people receiving messages from AI agents isn't scalable in the same way that people sending messages from AI agents is.
The token usage is really interesting. I would imagine the most efficient thing is to keep the state of everything persisted, and past the cache expiration window, to automatically start a new session with the previously persisted state instead of just a long running conversation.
If someone solves this part of continual effective compaction + selective resetting at cache expiry, they're going to make a ton of money. Right now, only the token insensitive can use these sweet features.
Are you all comfortable with the idea of agents running non stop with access to all your accounts? I would be so anxious all the time, thinking they would leak or delete my personal data, or get hijacked via prompt injection or a security vulnerability. I just cannot imagine using such a software without overview
Prompt injection is my biggest fear. Imho it is almost impossible to make a sort of tool that would be successful in detecting an injection - but maybe some antivirus/antimalware producers work on it…
The rest can be solved - separate account for agent spending (soon offered by your bank or neobank), separate email for agents (already exist) etc.
Sounds like “according John McAfee the threat of malware has been largely solved”.
Edit: it's way worse than that: the actual figures says that Mythos “only” falls for prompt injection 2.6% of the times. Maybe for an ML engineer used to work with unreliable tools that sounds impressive, but for security purpose having a system that fails every 40 attempts is outright catastrophic!
And we're just talking about known kind of prompt injection that are tracked by benchmarks, not any kind of 0-day vulnerability found by clever attackers.
"largely solved" as in they the models they trained don't fall for prompt injections as often but not "largely solved" as in the underlying issue is solved at all.
From FAQ: How is Grok Bot different from AI assistants? Bots have their own computer, so they can work inside your apps and tools. They also run in parallel, 24/7, even when your laptop is closed.
How does it work with login-walled sites like LinkedIn then? And what does "own computer" mean? X provisions a "private cloud" a'la Apple for your Bot?
The very first thing in their demo shows Grok logging in with the user's username and password to a website, presumably so it can perform actions and the human can get the blame for them. Apparently this is marketing and not terrifying to people.
Neat. I haven't had access to Grok Bot but I rolled my own async agent a while ago using systemd, Codex and old school UNIX stuff like sendmail/postfix/Maildirs. The model gets a rolling and continuously compacted context window in which mail arriving or timed wakeups cause Codex to resume a session with a prompt telling it to do stuff.
As it's linked to an SMTP gateway I interact with it entirely by email. It also has API keys for my internal bug tracker, it has its own git forks of my repos and so on.
Over time it's improved itself so it has adaptive wakeups, a nice monitoring dashboard so I can see what it's doing if I want, a bunch of skills, a nightly wakeup to reorganize its home directories and so on. It keeps a private knowledge base in its home dir and writes its own skills, and it has a persistent todo list. When the todo list is populated it wakes up more frequently so it can break work up into chunks, a kind of basic Ralph loop.
I've found this email-centric paradigm works pretty nicely for me compared to just running a local CLI agent. For instance, I can assign a bug to "Axiom" (it named itself) in YouTrack and the email notification causes it to wake up and start work on the bug. If it gets blocked, it uses its API key to post a comment back to the ticket. If I think of something I just quickly email it, so I'm using Gmail instead of dealing with finicky ChatGPT style UIs that don't have good support for labelling, compact UI and other nice things.
On older models its judgement about what to work on wasn't that good, but with the latest GPTs it's doing much better. I just got back from holiday and it had triaged and resolved over 25 different bugs, a nice selection too.
Because the whole thing runs in a dedicated UNIX user account it is able to modify its own configuration, which has helped as I developed this setup.
However, based on the blog post, Grok Bot sounds a fair bit more sophisticated. I haven't felt a need for more than one bot as I can't easily saturate even a single bot.
Cost: just uses my regular ChatGPT Plus subscription, $20/month, plus of course the server it runs on (which is shared with lots of other stuff). I'm not working with it heavily, perhaps if I did I'd need the $200/month sub.
That's what everybody is excited about; being able to write whatever they want. They just say 'Make a clone of Foo' and hey presto! The next step is to share your fabulous creation with the world. Lo, and behold, it looks just like Foo!
This is an interesting idea although I'm not sure how many companies are comfortable with giving SpaceXAI access to all your files and data. Outside of America this is, most likely, not going to fly.
Interesting. Unfortunately Musk's personal brand is so poisonous that I would never let him anywhere near my data. I'm curious if big businesses will have similar concerns and avoid tools from SpaceXAI regardless of how they are? I guess lots are already using it by default since the Cursor acquisition.
Musk burned billions in brand value by renaming Twitter, and he'll be doing the same with Cursor. The Cursor brand was the only wedge he had with enterprise customers, who seem to avoid Grok products altogether
If you have the time and are willing to do it, go check his Twitter profile. You can agree with him, but you will get it why some companies would avoid his products
Somehow the American AI industry managed to create a product I trust less than existing commercial offerings. Kudos, that is actually pretty impressive. Considering the stakes, one might also call it foolish. But I'm sure they've done their market research, they've identified their product niche and they're racing to meet it at the bottom.
Although I'll probably never use Grok, this is an interesting release. I always thought GPT work and Claude Cowork are a bit awkward in their positioning. I'm still not sure what the real difference is with standard chat. This looks like the same thing with seamless memory and skills creation.
This seems like the middle ground between work mode and OpenClaw/Hermes (which I've never tried, too scared to mess something up). I think the humanization of the agents is cute and makes sense UX wise. I hope to see alternatives soon.
Biggest advantage is each one owns its own routines, context, and domain, and they can communicate between each other. Similar to hermes they build out their own skills, but by keeping the bots separated by domains, you end up getting better results out of them.
Additionally though each one has their own computer, which means async work feels like it actually works. I haven't had to juggle worktrees for the last month.
Biggest downsides are token expenditure. I've used more tokens this month than not this month. That's not a typo - I've used less tokens in the last 5 years prior to this month than I have this month. Always on perpetual agents use a LOT of tokens. IMO this is building for the future state where tokens are vastly cheaper, ie in a post-ASIC world.
The coolest thing I had it do for me was sourcing fabric for swag: https://image.non.io/d83664c1-5807-4a18-abe4-41928c198410.we...
I wanted to make something that didn't feel like just my logo on a shirt, so I had one of my bots reach out to ~40 fabric suppliers in vietnam, negotiate prices, lock one in, and get samples made. First samples should be finished today. It's been something I've wanted to do for ages, so it was cool seeing it actually happen. The fabric supplier bot worked with one of my prototyper bots to create a randomly generated pattern using my logo, which it then sent as a .ai file to the supplier.
Isn't this one of the problems foreseen with this? For you, it was a single prompt - for 40 companies, this probably took up some time.
What happens when fifty people fire off a 15-second "get me a shirt" prompt? When five hundred, five thousand, five million do?
If your business is selling tokens, it'd be extremely lucrative for you if the whole society relies on tokens to perform basic operations. That's where we're heading to.
One of the difficulties of sourcing this is a lot of the suppliers in vietnam are only contactable via whatsapp. Emails are monitored far less. It's one of the reasons I haven't been successful with this in the past despite trying - it's a very word-of-mouth network.
If someone solves this part of continual effective compaction + selective resetting at cache expiry, they're going to make a ton of money. Right now, only the token insensitive can use these sweet features.
The rest can be solved - separate account for agent spending (soon offered by your bank or neobank), separate email for agents (already exist) etc.
[1]: https://x.com/bcherny/status/2086520950259118464
Edit: it's way worse than that: the actual figures says that Mythos “only” falls for prompt injection 2.6% of the times. Maybe for an ML engineer used to work with unreliable tools that sounds impressive, but for security purpose having a system that fails every 40 attempts is outright catastrophic!
And we're just talking about known kind of prompt injection that are tracked by benchmarks, not any kind of 0-day vulnerability found by clever attackers.
The core is there. But there is some work to be done to have a nicer shell and all, which I’m currently focusing on.
How does it work with login-walled sites like LinkedIn then? And what does "own computer" mean? X provisions a "private cloud" a'la Apple for your Bot?
https://image.non.io/4022ec77-be07-4baa-97e8-ad9d8d9aeb8a.we...
After you do you just tell the bot you're done logging in and it'll keep driving. And yea, it's a separate VM for each bot.
Source: had access for the last few weeks.
As it's linked to an SMTP gateway I interact with it entirely by email. It also has API keys for my internal bug tracker, it has its own git forks of my repos and so on.
Over time it's improved itself so it has adaptive wakeups, a nice monitoring dashboard so I can see what it's doing if I want, a bunch of skills, a nightly wakeup to reorganize its home directories and so on. It keeps a private knowledge base in its home dir and writes its own skills, and it has a persistent todo list. When the todo list is populated it wakes up more frequently so it can break work up into chunks, a kind of basic Ralph loop.
I've found this email-centric paradigm works pretty nicely for me compared to just running a local CLI agent. For instance, I can assign a bug to "Axiom" (it named itself) in YouTrack and the email notification causes it to wake up and start work on the bug. If it gets blocked, it uses its API key to post a comment back to the ticket. If I think of something I just quickly email it, so I'm using Gmail instead of dealing with finicky ChatGPT style UIs that don't have good support for labelling, compact UI and other nice things.
On older models its judgement about what to work on wasn't that good, but with the latest GPTs it's doing much better. I just got back from holiday and it had triaged and resolved over 25 different bugs, a nice selection too.
Because the whole thing runs in a dedicated UNIX user account it is able to modify its own configuration, which has helped as I developed this setup.
However, based on the blog post, Grok Bot sounds a fair bit more sophisticated. I haven't felt a need for more than one bot as I can't easily saturate even a single bot.
Cost: just uses my regular ChatGPT Plus subscription, $20/month, plus of course the server it runs on (which is shared with lots of other stuff). I'm not working with it heavily, perhaps if I did I'd need the $200/month sub.
I wonder if they're going to try and compete against Slack with X.com chat?
This is an interesting idea although I'm not sure how many companies are comfortable with giving SpaceXAI access to all your files and data. Outside of America this is, most likely, not going to fly.
with Grok, that's a tremendous assumption
Ah, you mean guardrails for _your_ protection...!
Dang - might be worth merging these two.
GitHub gives 404 after logging in so I can’t event try it.
Although I'll probably never use Grok, this is an interesting release. I always thought GPT work and Claude Cowork are a bit awkward in their positioning. I'm still not sure what the real difference is with standard chat. This looks like the same thing with seamless memory and skills creation.
This seems like the middle ground between work mode and OpenClaw/Hermes (which I've never tried, too scared to mess something up). I think the humanization of the agents is cute and makes sense UX wise. I hope to see alternatives soon.