Michal Pandy

Co-founder @medi_search

[Michal Pandy @michal_pandy]

Have a new dataset and not sure which pre-trained model to fine-tune?

We present GBC, a scalable approach for selecting the ideal source models for fine-tuning (without actually fine-tuning!). To appear in #CVPR2022.

Paper: https://t.co/Fbm1ESPM0T

1/4 🧵

[Michal Pandy @michal_pandy]

@giffmana @KanishkaNarayan @AnthropicAI Arm?

[Michal Pandy @michal_pandy]

@jrarribas @EvidenceOpen https://t.co/UyBy1lrLqC es mejor :)

[Michal Pandy @michal_pandy]

@loaibassam @yourealazyfvck You get the same email from Apple, but it goes from 15% to 30%.

[Michal Pandy @michal_pandy]

@pmddomingos gmail at google and playstation at sony were also side projects :)

[Michal Pandy @michal_pandy]

@RoupenMD Original is often whoever wins distribution first. Two teams can start at the same time, but the one with massive marketing becomes the baseline. Everyone else looks like a follow-on when they finally reach users.

[Michal Pandy @michal_pandy]

@zxchyanez @ycombinator @medi_search one way to find out: https://t.co/MLRMeuOfQT :)

[Michal Pandy @michal_pandy]

@AlleyBo55 @ycombinator @medi_search https://t.co/kQajw60JOc

[Michal Pandy @michal_pandy]

@gritfollow @ycombinator @medi_search Hi! We did run the benchmark using Gemini 3. We could only get it to 13.3 with maxed out reasoning. However, we did not want to share this in the original chart because this is not an official result from their published evals. The OpenAI results are here: https://t.co/wzB8D4ijnQ

[Michal Pandy @michal_pandy]

https://t.co/fzaEu1AxEs

[Michal Pandy @michal_pandy]

We’ve just released our most advanced medical AI yet: MediSearch Max.

We tested it on HealthBench Hard, a new benchmark from OpenAI designed to test models on difficult medical cases. Max scores 53.3%, beating both GPT-5 and GPT-4o.

Give it a try. Link in the comments. 🙂

[Michal Pandy @michal_pandy]

@GladiaLab What happens under Adam, ReLU, RoPE, or quantization? collisions appear?

[Michal Pandy @michal_pandy]

@DrKesbeh I am not super familiar with their quality level. We have some comparisons based on the performance they published: https://t.co/zfLMDlvBPV. Happy to hear your thoughts once you try MediSearch though 🙂

[Michal Pandy @michal_pandy]

@DrKesbeh You may enjoy https://t.co/eXD4lzKEow, we index all new medical literature every day 🙂

[Michal Pandy @michal_pandy]

@thesamparr medisearch

[Michal Pandy @michal_pandy]

@giffmana @cloneofsimo Bayesians: it’s just the chain rule of probability btw

[Michal Pandy @michal_pandy]

@AravSrinivas Google - OK
Perplexity - Better
MediSearch - Best

[Michal Pandy @michal_pandy]

@agihippo It is a bit unfair to judge science this way. Someone can riff on it and _maybe_ make something that will work better than MLPs. That's how progress happens.

[Michal Pandy @michal_pandy]

@isaachasson We need meta-consultants to help the consultants. Somebody should start McMcKinsey.

[Michal Pandy @michal_pandy]

@GillVerd @ylecun I found training an EBM extremely challenging for high dimensional data using non-smooth architectures. Always had to make the nets residual, Lipschitz etc. + other tricks. How do you plan to overcome this @GillVerd?

[Michal Pandy @michal_pandy]

@musho @copy_ai Amazing, congrats and all the best