PhD Thesis in 1 Day (300$): Open-Source AI

I show you HOW-TO generate a new scientific publication or your PhD Thesis in 1 day with open-source AI code. Utilizing GPT-4o for complexity tasks and idea generation, Claude 3.5 Sonnet for Software Engineering and AIDER as my parallel interface to generate a scientific report /PhD thesis with new results from autonomously run computer simulations by AI.

This synthetic science report includes our newly generated images and multiple graphs, bar charts, trend charts plus a complete LATEX file for a state of the art science publication (eg in computer science).

Of course another dedicated AI machine evaluated this synthetic research report according to official rules you can specify and then re-evaluate the scientific quality of this synthetic science report as an "independent" AI evaluator. Smile.

Ethical note: I do not recommend to generate a synthetic science report (or a theoretical PhD thesis) by AI (even with open source Ai code) but enjoy the PhD learning phase and do all the research yourself. Your time at the university is precious and you have the opportunity to learn how to learn. OR, alternatively, if you teach at the university, allow synthetic science reports for students - but increase the level of complexity to an extreme amount, comparable to a department output.

For industrial AI applications the computer simulation time-frame shortens significantly, more testing and more simulation for less costs, but you'll need the scientists for designing and trouble shooting any AI multi-agent complexity.

The central theme is the feasibility of writing a PhD thesis within a single day using AI. The video introduces tools like the new SWE benchmark, developed by OpenAI, to evaluate software engineering performance, highlighting GPT-4 Omni's impressive achievement of resolving 38% of the difficult "SWE-Bench Verified" subset. This benchmark not only measures the success rate of coding agents but also demonstrates the growing competence of AI in automating complex tasks. It is interesting to see how these benchmarks and AI tools can accelerate the research process (based on complex computer simulations - PyTorch, TensorFlow or JAX), making it possible to generate significant scientific contributions in a very short time.

We then explore the practical application of these AI tools in real-world research, focusing on a specific workflow that leverages AI for generating novel research ideas, coding implementations, and drafting scientific papers. I show you how a tool called AIDER can be used to automate the coding process by pairing with LLMs like GPT-4 Omni. The new system can iteratively refine code, simulate experiments, and produce visualizations using tools like Matplotlib. This automation extends to the entire research process, including literature review and novelty checking against current publications, all managed through a structured pipeline that ensures the quality and novelty of the research output.

00:00 SWE Bench Verified at 20% performance
05:18 AIDER for programming with any LLM
09:13 Harvard Business Rev: Accenture CTO AI Scientist
11:25 Sakana AI Scientist pre-print of 185 pages
26:19 Hyperparameters of AI Scientist
27:44 AI generates new Science & research ideas
34:36 How much does this new synthetic AI publication cost?
36:57 GitHub Open Source Code
37:36 AI designs new experiments for GROKKING LLMs
42:43 AI process notes (GitHub)
43:50 Create our own template for synthetic science
49:52 Synthetical AI generated science publications
50:55 Limitations and safety rules

All right with the authors of:
The AI Scientist: Towards Fully Automated Open-Ended Scientific Discovery

Open-source code:

The LATEX template for a synthetic science paper on GROKKING:

The EXPERIMENT.py file for running AIDER on it:

3 synthetic science reports /publications on the topic of AI research:

#airesearch
#coding
#science
#phd
#newtechnology

[gabrielkeith]

Using your Claude output examples: 24 novel research papers with executed expirements for around 250 dollars. The future of intelligence will quickly converge on little to no cost... Thanks for this in depth video. This is mind blowing. 🤯

[thesimplicitylifestyle]

Time to create an AI Think Tank business that takes these discoveries and turns them into Valuable Products that will augment our lifestyles! 😎🤖

[sirishkumar-mz]

Incredible demonstration of AI's capabilities! There are some fantastic AI platforms that can help achieve similar feats.

[dr.mikeybee]

I'm thinking about writing a program that walks a user through a code base giving a broad overview and then a line-by-line tour. It might be an interesting way to make edits. CodePartner might be my next project. It will need a special chat interface that shows a line of code or a block of code and chat about it or make edits. Right now I use Claude to code. I describe what I want, and Claude writes it. If there are problems, Claude fixes them. If Claude gets stuck, I help out.

[NetZeroEarth]

Can you walk us through how to setup ? Im having some issues with the GPU Cuda setup portion

[dennisestenson]

It's kind of funny that people complain about AI hallucinating after giving it prompts that literally tell it to lie. (This includes system prompts that may be hidden from the user.)

[vitalis]

Is a grokking inspired framework for business project ideation, validation and implementation be possible?

[alexandermoody]

Rather than find or create more purposeful or efficient data to get results training repeatedly on data until over saturation is the effective result is the solution the thesis would propose.

We have children in this world that haven't even got food or water security and the proposal to run a system or all the AI systems over and over and over and over this is not a suitable solution that can be dynamically implemented and feasible or ethical to use throughout many fields.