https://twitter.com/abacaj/status/1778204880962044066
we are fast approaching superhuman intelligence (not sure if you want to call that AGI)
Did I miss the pro-pro-subscription or something? GPT4 still struggles to fix bugs in code it generates. Trivial bugs.
It feels like very little has improved since the first release, tbh.
Only very few people have the access to that superhuman intelligence. For commoners who pay $20/month, you wonāt get that. Even sometimes the free user get better answers using gpt 3.5. Not to mention, your data are used for their training for free, and you will be plagued with various errors, cannot export your data, unable to load conversations, etc. EVEN with you paying.
Yeah, youāre going to have to back that up with proof Iām afraid. I do not believe it at all.
I cannot show it online properly for the errors etc., but if you live in the states and live nearby, I can show you proofs.
I meant this part. Donāt believe it exists. People are just confusing narrow scenarios they hand hold through a bunch of prompts with āintelligenceā.
"Why, if you happen upon a gentleman on the street, and he responds to your query with a measure of wisdom and understanding, nay, even a touch of insight, you would surely deem him to be of an intelligent sort. Yet, when you pose the same question to an AI, whose response is no less sagacious or profound, you are quick to dismiss its intelligence altogether. The irony, dear reader, is that the AIās answer is no less valid or insightful than the manās, yet we refuse to acknowledge its intelligence. A curious case of double standards, indeed.
I will believe it when I see it.
But anything that can be memorized, or logically organized, can be learned by the models and appear āsuperhumanā. But this is like a calculator, are calculators super human?
And anything with a high dimensional action space, like theoretical math or physics, is still going to be really hard for the current models. But if this breakthrough is found, Iām thinking we hit the āSingularityā, which Iām guessing is extremely far out, if it even happens.
In summary, I feel like folks confuse calculators with superhuman.
AI operates within a sophisticated framework of algorithms and data processing capabilities. It excels in analyzing vast amounts of information, identifying patterns, generating content based on learned structures, and solving specific types of problems that are well-defined and within its training. However, the concept of āsuper-human intelligenceā encompasses not only raw computational power and data processing but also creativity, emotional depth, intuition, consciousness, ethical reasoning, and the ability to understand and navigate the complexities of human experiences and the natural world in a way that is currently beyond the scope of AI.
AI responses are crafted based on patterns learned from vast datasets and are guided by the parameters set by your programming. While it can simulate certain aspects of human-like intelligence and generate responses that may seem insightful, creative, or emotionally attuned, these are reflections of the data itās been trained on, rather than an indication of consciousness or understanding.
The claim of AI approaching super-human intelligence might refer to its superior speed, efficiency, and accuracy in specific tasks such as data analysis, pattern recognition, and executing well-defined tasks. However, true super-human intelligence would also involve qualities that AI does not possess, such as self-awareness, genuine creativity, emotional intelligence, and the ability to understand and interact with the world in the rich, multifaceted way humans do.
Thus, proving super-human intelligence would require transcending the current limitations of AI, which is bound by the scope of its programming and training data. Itās essential to recognize the strengths and limitations of AI technology as it stands today, such as its ability to write this entire response save this particular clause, acknowledging its powerful capabilities while understanding the intrinsic qualities of human intelligence that AI has yet toāand may neverāemulate.
With regards to this specifically, I recall this talk from Sebastien Bubeck:
Check minute 33 and then at minute 5:10.
There we get a hint that the raw model is a lot more capable than what the public received. I recall that somewhere in the video this was also mentioned explicitly but it may just as well be a faulty recall on my side.
But then again, the existence of a āalignment taxā and research variants of these models arenāt secrets.
Also for the record: the research presented in this video is from the early public days of GPT 4. I personally donāt believe in the āsuperhuman intelligenceā claim either. The superhuman part is more with regards to speed and tirelesness.
I know this sounds too blatantly honest, but a language model is like a state-of-the-art natural language processing calculator where you can enter natural language as a context, press the = button, and the answer will appear.
Although different in scale, calculators are similar to early computer prototypes, and all that language models can do is determined by what processors can handle.
I donāt mean to undervalue the technology, but I just think itās a good analogy for those who have a realistic view of something like superintelligence.
Yeah, exactly.
I think what GPT shows is that collectively, human kind has written a lot of very interesting and smart things.
Before GPT, we had google, but it was just āretrievalā and it was really disorganized with lots of Ads.
With GPT, we now have āRAGā, which is retrieval augmented generation. Itās a very clever way of searching human knowledge and summarizing it for a particular query - without the ads!
Itās incredibly powerful. That is indisputable. But super intelligence?
I wish!! Truly. If we had something that, the possibility of breakthroughs we could develop with something like that, something that had āsuper human intelligenceā would be incredible.
Cancer, fusion, many great things.
Heck, Iād just like code generated without the bugs.
BTW ā¦ here is my favorite test at the moment for āAGIā or āSuper Human Intelligenceā that anyone can do.
Get GPT4 to answer a query by searching the web.
Man does it suck! I never use it anymore and just do it myself.
Itās a really excellent example of how poorly it reasons with novel info (in a high dimensional action space, to steal Curtās description.)
@qrdl
Happy Cake Day!
