Everyone doing the research into LLMs actually, that's sort of the race right now. The race is towards 'reasoning'.
Perhaps another way to look at LLMs is that we're AI to understand language. Right now, that language is English. So as we teach it English, and it understands English better, we can provide it the knowledge base needed for reasoning. So a simple example might be, here's the rule book for a boardgame. If the rules are written well enough, we would expect that the LLM would be able to reference this rulebook and others, and perhaps appendices, to be able to official rules on the boardgame without assistance.
And that should also mean, based on it's experience in the knowledge reading, it would be able to infer, exception cases as well.
You can draw a straight line from there, to apply it to Law.
And from there, you can start to apply it to PHD level problems. If we provide AI the knowledge base to answer certain questions, how long it would it take to come up with the same response as a PHD person would researching the same topics. And if it takes for instance, 2 years, for a PHD person to read through all of that knowledge, and come up with an experiement to solve a problem. It could take LLMs say a week, or less to rifle through all that knowledge and be able to come up with the same experiments to test.
So one could draw a line from that, and say, okay, what if we give it all the knowledge, from all the fields, related or not, and with access to all the worlds knowledge and experiements, could it now solve or at least come up with some potential solutions or provide breakthroughs in our thinking, to solve issues like
climate change,
fusion reactors
materials engineering
etc.
so as long as we continue to improve LLMs ability to understand language, context, and math, etc, then we have a way now to process huge amounts of data into a single entity, absorb and understand the specialities that normally we hire each individual specialist to do, and now it' can use all of these connected knowledge to develop potentially solutions we never knew were connected.
We are now moving to the world of AI, that's why I stopped Googling. you'll never get the answers you're really looking for. But Perplexity is probably the best web based AI. It's fast an useful.
*****
can chatgpt o1 solve PHD level problems
******Yes, ChatGPT o1-preview has demonstrated the ability to solve PhD-level problems across various disciplines, particularly in physics, chemistry, and biology[1][2]. In a notable test, physicist Kyle Kabasares challenged the model with complex problems from Jackson's Classical Electrodynamics, a notoriously difficult textbook used in advanced physics courses[1].
The o1-preview model showed impressive capabilities:
1. It solved a problem from Jackson's textbook in just under two minutes, a task that typically takes graduate students about a week and a half[1].
2. The model tackled a complex two-part problem involving electromagnetic wave scattering, demonstrating human-like reasoning and problem-solving skills[1].
3. OpenAI reports that the o1 models perform similarly to PhD students when working on physics, chemistry, and biology problems[2].
4. The o1 models scored 83% on a qualifying exam for the International Mathematics Olympiad, a significant improvement over the previous GPT-4o model, which only solved 13% of problems correctly[2].
It's important to note that while the o1-preview model has shown remarkable abilities, it is still in its early stages. OpenAI is taking a cautious approach to releasing these models, with limited access currently available to ChatGPT Plus or Teams account holders[2].
Citations:
[1] https://www.webpronews.com/chatgpt-...ics-has-ai-mastered-advanced-problem-solving/
[2] https://www.fastcompany.com/91189817/openais-new-o1-models-push-ai-to-phd-level-intelligence
[3] [4]
[7] https://community.openai.com/t/open...to-have-super-iq-phd-level-competence/1095182
[8] https://www.linkedin.com/posts/robi...my-phd-code-activity-7241816397138272256-ez7S
