🖖 Rik Doclo 💯’s Post

💥💥💥 New GPT-4o model outperforms humans on MathVista for the first time! 📊 Scores: Human avg: 60.3 GPT-4o: 63.8 MathVista: mathvista.github.io #machinelearning

Classic example of GPT4 giving you a wrong answer while sounding so sure of it! The right answer to my question below is YES. Streamlit will add the key associated with the button to the session_state.

GraphRAG is RAG for riches It cost me approximately $5 to create a simple, functioning example using only ONE text file! Used: gpt-4o & text-embedding-ada-002 Next step: try to achieve the same response accuracy with Llama and other open-source models.

GPT-4o, a mini GPT5? Having played with most models in the world over the last 24 months, I was pleasantly surprised by GPT-4o for a few reasons. Firstly, GPT-4o’s demos during the spring release video were honest and real-time. Yes, they were inevitably cherry-picked, but they were really slick, and I love that they didn’t over-edit the video, showing moments like when 4o thought one of the researchers was a table…😂. This shows they have confidence in the robustness of the technology. Even under pressure cooker situations. Secondly, the zero-shot performance seems to be off the charts. This is something few are shouting about, but I think it represents an unbelievable step forward in intelligence and reasoning. Still NOT AGI but a step forward over GPT4t And thirdly, its speed. 4o is getting a bit of a battering on X etc. at the moment, with some saying speed = lower intelligence, but with speed, the use cases where agents can be deployed are expanding, and if I can do 3 agents in the time it took GPT4 to produce one response… it’s going to be orders of magnitude more useful in more use cases. My takeaway 🥡 The combination of speed and improved intelligence could leave everything that came before in the dust. As close to AGI as we currently have. Proof read by GPT4o #gpt4o

Closed models and open models are becoming increasingly similar on MMLU. There are now far more important benchmarks, such as HumanEval or SWE. Nevertheless, there is currently a trend that the previously large gaps are no longer being maintained. And as long as GPT-5 does not again show similar jumps in all benchmarks as GPT-4 (shown below), further convergence is to be expected.

GPT-4o mini is my new default LLM: It costs a fraction of GPT4, but comes with the same vision capabilities (image understanding) and same 128k context length. The performance is only slightly worse compared to the bigger models: https://lnkd.in/ebEJsDkn

Absolutely crazy that I'm able to run Phi-3-medium locally using DirectML. It's a pretty large model with 14B parameters and it's running stable. The reasoning skills of this model however, well... see for yourself. It does run simple questions fine and the instruct capabilities are promising. It's just bonkers that we thought we wouldn't be able to run GPT-like models locally, but a year later we're running models that are more capable than GPT3.5 on a simple desktop.

#30daymapchallenge As an old-school (former) ArcView 3.2 user, I find it fascinating that you can now generate a map by just asking a GPT model. The result might be something to consider, especially as it might get better in the coming days. My question to the GPT was, "Generate a map with a new projection with Bangui at the center."

This is GPT o1 being scored on a MENSA level IQ test, and compared to older models. Thanks to Maxim Lott , full article here : https://lnkd.in/gW5YXzcE

Karpathy’s Zero to Hero series ——————————————— Build GPT2 (124M) from zero to end The video start with empty file and end up with a GPT-2 (124M) model: - first it builds the GPT-2 network - then it optimizes it to train very fast - then it sets up the training run optimization and hyperparameters by referencing GPT-2 and GPT-3 papers - then it brings up model evaluation, and - then run actual training overnight - Finally it looks through the results. The "overnight" run even gets very close to the GPT-3 (124M) model. Video: https://lnkd.in/d4w4z-td GitHub repo: https://lnkd.in/dhu6H5Dd