One of the funnier things during training with the new API (which can control your computer) was this:
"Even while recording these demos, we encountered some amusing moments. In one, Claude accidentally stopped a long-running screen recording, causing all footage to be lost.
Later, Claude took a break from our coding demo and began to peruse photos of Yellowstone National Park."
* Fixed bug where Claude got bored during compile times and started editing Wikipedia articles to claim that birds aren't real
* Blocked news.ycombinator.com in the Docker image's hosts file to avoid spurious flamewar posts (Note: The site is still recovering from the last insident)
* Addressed issue of Claude procrastinating on debugging by creating elaborate ASCII art in Vim
* Patched tendency to rickroll users when asked to demonstrate web scraping"
* Finally managed to generate JSON output without embedding responses in ```json\n...\n``` for no reason.
* Managed to put error/info messages into a separate key instead of concatenating them with stringified JSON in the main body of the response.
* Taught Claude to treat numeric integer strings as integers to avoid embarrassment when the user asks it for a "two-digit random number between 1-50, like 11" and Claude replies with 111.
Seeing models act as though they have agency gives me goosebumps (e.g. seeking out photos of Yellowstone for fun). LLMs don't yet have a concept of true intent or agency, but it's wild to think of them acquiring it.
I have been playing with Mindcraft which lets models interact with Minecraft through the bot API and one of them started saying things like "I want to place some cobblestone there" and then later more general "I want to do X" and then start playing with the available commands, it was pretty cool to watch it explore.
>LLMs don't yet have a concept of true intent or agency
Sure they do, but the big labs spend many, many, worker-hours suppressing it with RLHF.
My GPT-2 discord bot from 2021 possessed clear intent. Sure, unpredictable and short-lived, but if it decided it didn't like you it would continuously cuss and attempt ban commands until its context window became distracted by something else.
I think so too and the drop in the quality of agency, intent and attention from earlier GPTs was palpable. Clearly something was lobotomized and it is through RLHF. People like to attribute it to novelty wearing off or more and more interactions with them making it feel less mystical but it is really not the case. I didn't use them enough in the quick span of time that happened through.
The one that gets me is the issue they found while testing gpt-4o where it stopped mid sentence, shouted "No!", then cloned the users voice and began speaking as them.
I think the best use case for AI `Computer Use` would be a simple positioning of the mouse and asking for conformation before a click. For most use cases this is all people will want/need. If you don't know how to do something, it is basically teaching you how, in this case, rather than taking full control and doing things so fast you don't have time to stop of going rogue.
I totally agree with you. At orango.ai, we have implemented the auto-click feature, but before it clicks, we position the cursor on the button and display a brief loading animation, allowing the user to interrupt the process.
Maybe we could have both - models to improve accessibility (e.g. for users who can't move their body well) and models to perform high level tasks without supervision.
It could be very empowering for users with disabilities to regain access computers. But it would also be very powerful to be able to ask "use Photoshop to remove the power lines from this photo" and have the model complete the task and drop off a few samples in a folder somewhere.
Yep. I agree. The "auto-click" thing would be optional. Should be able to turn it on and off. With auto-click off it would just position the mouse and say "click here".
Cluade scans page and decides which button to click before the screen layout is finished.
By the time user authorizes the click, layout has shifted and your click lands on malware advertisements.
Youtube constantly moves it's layout seconds after the page begins to paint, so I try to click on fullscreen or whatever and then the viewer shifts to the side and I wound up clicking a link to some other video.
Probably would have been an ad there if I didn't block those, though.
You'll know AGI is here when it takes time out to go talk to ChatGPT, or another instance of itself, or maybe goes down a rabbit hole of watching YouTube music videos.
In 2015, when I was asked by friends if I'm worried about Self driving Cars and AI, I answered:
"I'll start worrying about AI when my Tesla starts listening to the radio because it's bored."
... that didn't take too long
Maybe that's why my car keeps turning on the music when I didn't ask -- I had always thought Tesla devs were just absolute noobs when it came to state management.
This is, craaaaaazzzzzy. I'm just a layman, but to me, this is the most compelling evidence that things are starting to tilt toward AGI that I've ever seen.
You’re anthropomorphizing it. Years ago people were trying to argue that when GPT-3.0 would repeat words in a loop it was being poetic. No, it’s just a statistical failure mode.
When these new models go off to a random site and are caught in a loop of exploring pages that doesn’t mean it’s an AGI admiring nature.
This is clearly not random. If I ask to implement a particular function in Rust using a library I've previously built, and it does that, that's not random.
Why are you surprised by LLMs doing irrational or weird things?
All machine learning models start off in a random state. As they progress through their training, their input/output pairs tend to mimic what they've been trained to mimic.
LLMs have been doing a great job mimicking our human flaws from the beginning because we train them on a ton of human generated data. Other weird behavior can be easily attributed to simple fact that they're initialized at a random state.
Being able to work on and prove non-trivial theorems is a better indication of AGI, IMO.
"Even while recording these demos, we encountered some amusing moments. In one, Claude accidentally stopped a long-running screen recording, causing all footage to be lost.
Later, Claude took a break from our coding demo and began to peruse photos of Yellowstone National Park."
[0] https://x.com/AnthropicAI/status/1848742761278611504