Today, in 2023, it's good to remember that most of the capabilities that the tech industry assumes LLMs to already possess aren't yet within reach. Tread this space carefully, and beware of shiny demos. Last year, I was repeatedly told that the upcoming GPT-4 was already AGI.
@fchollet
-
Domestic Robotics’ Unfulfilled Promises Before 2020
By
–
Around that time, most people believed that domestic robotics would be a solved problem before 2020 (how could it not? did you see that Boston Dynamics video?), which led to a huge wave of investment in robotics — which of course didn't pan out.
-
NIPS 2016 AGI Predictions on Reinforcement Learning
By
–
Half of my conversations at NIPS 2016 were about how deep RL trained on game environments and infinite simulations would lead to AGI in 5-10 years (this was immediately post-Alpha Go).
-
Human-level language understanding remains years away despite progress
By
–
In 2016, when I tweeted that human-level language understanding was many years away (which is still the case now, though we're closer), mind the context: this was in response to many people, including prominent VCs, claiming that then-current AI was nearly there and was about to
-
AI Progress: Applications vs Generality and Future Capabilities
By
–
Remember — we are making progress on AI (though far more on applications than on generality, which remains largely a green field). The progress is significant in speed and magnitude. But the conventional wisdom of the tech community about current and near-future AI capabilities
-
AI capability predictions: 2016 expectations versus 2023 reality
By
–
In 2016, people didn't anticipate next-token prediction models (which were LSTMs then) to be this capable, but they did expect AI to soon have most of the capabilities they show now. Remember the chatbot mini-bubble of 2017? It was predicated on 2023 capabilities coming circa
-
AI Predictions: Optimistic Bias and Historical Timeline Assessments
By
–
If I rate my own AI predictions, every wrong one was on the optimistic side (e.g. in 2018 my timeline was mass-deployment of self-driving was 2023, which sounded extremely pessimistic back then). And people think I'm a skeptic (which I am not — historically I've been slightly
-
Self-driving cars progress stalled since 2016 predictions
By
–
If you told people in 2016 that in mid-2023 there would no way for a member of the general public to take a driverless ride in Silicon Valley, *no one* would have believed you. After all, self-driving cars were already on the streets and were making insanely fast progress, right?
-
Self-Driving Cars: 2016 Predictions vs Reality Debate
By
–
Funnily, there seems be as many folks saying "no one ever claimed in 2016 that self-driving car were anywhere close to ready" (fact check: it was the conventional wisdom in the tech crowd back then that they'd be mass-deployed before 2020) as folks saying "self-driving car were
-
Spotting Students Who Memorize Without Understanding: The AI Analogy
By
–
This is also how you spot students who pass their exams purely by memorizing tons of previous tests without understanding the class materials (a pretty good analogy for AI): you test them on something they didn't expect, that requires thinking from first principles.