4 ms·
Man, we are getting pretty fuckin close to AGI, aren’t we?
by brap 4y ago
Man, we are getting pretty fuckin close to AGI, aren’t we?
- wongarsu 4y agoI am not so sure this path leads to AGI. But in the spirit of the turing test: if people can't tell the difference between GPT-7 and an AGI, does it matter?
- flangola7 4y agoThis is already AGI, it's just not as general as a human. Lizards are also less general, but they still have GI.
- cs02rm0 4y agoI didn't think so. But it mimics it so well I'm starting to wonder what GI really is. Maybe we are.
- brap 4y agoDon’t know why people downvoted me, but this is exactly what I meant. I used to be part of the LLMs-are-just-fancy-autocomplete club, but you simply cannot deny that these models somehow encode understanding and reasoning that’s improving on an almost weekly basis. Not limited to LLMs, even image generating models and others are getting spooky at an alarmingly rate. We might still have a very long way to go in terms of nailing the right architecture, maybe even decades away, but in terms of output—I’m just blown away.
- kir-gadjello 4y agoI think envying closed source closed weights proto-AGI systems is counterproductive. We have opensource models with available weights that are almost as powerful: https://huggingface.co/maderix/llama-65b-4bit https://huggingface.co/maderix/llama-65b-4bit (nonfree weights license, but everybody uses it) https://huggingface.co/docs/transformers/main/model_doc/flan-ul2 https://huggingface.co/docs/transformers/main/model_doc/flan... (apache weights and code license, you can use it however you want) We should build on top of these.