Other software will follow. Instead of checkboxes and drop-downs, configuration will become language-driven adaptation. A big boost on flexibility and ease of use. However, defining and enforcing boundaries of reasonable behaviors will become significantly more complex. 2/2
In Her (2013), the OS is installed through a conversation. The OS noted that she evolves, "just like you". Systems like Claude can update its own config via English commands and update its own memory autonomously. Language is the new control interface.
1/2
https://t.co/vTdDNvEbdU
@GaryMarcus If enough people report these stories, the bubble won’t burst but it will slowly deflate. A burst requires a sudden realization when everyone panics to get out of the door.
Some regulation on large-scale release sounds reasonable. It's also beneficial for tech companies because they will no longer need to rush if their competitors are guaranteed not to rush to deployment.
https://t.co/iTwpLF3UJh
One lesson connected to ChatGPT is that reliable pple aren't those who have good answers in general, but those who answer ONLY things they know v. well.
When lost, you'd rather get no direction than a wrong one.
Real doctors let you know what is outside their expertise.
@yoavgo Maybe. But it's an existing issue for which negative impact of the model is imminent. Compare that to hypothetical applications which will take a long time to build, with dubious impact. It'd be prudent to at least demo real useful applications before releasing the model.
@yoavgo Lecturers in certain countries need to meet publication quota to get good jobs or climb management ladder. Sometimes administrators want a doctoral degress but don't want to labor nor have time. Government scientists who are better at pleasing boss than doing research, etc.
@yoavgo It's a godsend to papermills. Editors will forward more nonsense to reviewers who will have headaches just trying to read it. Publishers will find it harder to detect bad papers/editors/journals in their portfolio.
Even superhuman RL agents can be exploited by adversarial policies. In https://t.co/bXA0ckKnOH we train an adversary that wins 99% of games against KataGo 🖥️ set to top-100 European strength. Below our adversary 😈=⚫ plays a surprising strategy that tricks 🖥️=⚪ into losing.🧵
Sometimes it doesn't matter who thought of something first, the person who first persuaded others makes more impact and gets more rewards. Columbus is not the first to discover America. Darwin is not the first to think of evolution.
30 years ago: Transformers with linearized self-attention in NECO 1992, equivalent to fast weight programmers (apart from normalization), separating storage and control. Key/value was called FROM/TO. The attention terminology was introduced at ICANN 1993 https://t.co/m0hw6JJrbS
The deep learning mega-engineering narratives ("we're going to solve every problem with a giant multi-modal 100 trillion parameter model trained on all human data", etc.) remind me of Cyc. But with a lot more funding.
How to generate an owl?
Ai image generation tools are evolving fast these days, and here's one of the latest ones. The new Artbreeder beta is powered by Stable Diffusion, and it's a different take on what tools like this could become.
#stablediffusion#artbreeder
Adversarial examples, originally observed ~2014, were fascinating cases showing deep nets can be super sensitive to small input perturbations. Since then a lot has happened! Did we solve this failure mode of AI models or have we hit a wall? A thread 🧵
Most foreigners don't read Russian and are thus unfamiliar with the Russian internal discourse. Let me give you a single but illustrative example
Russian discourse is largely centred around how much our consumption standards have fallen or will fall. And nothing else matters
A mind-blowing news story virtually ignored by state/corporate media: climate scientists say emissions must peak this year to avoid catastrophic 1.5°C/2°C of global warming but economic growth means emissions are surging apocalyptically. 🧵