🏅 Top traders

sorry for slow resolution, stopped spending much time on here

@Mag, Can this resolve yes?

Phi-3 / Phi-3-mini (≈3.8 B parameters): Benchmarks show Phi-3-mini scores close to GPT-3.5 on academic tasks (e.g., MMLU, HellaSwag) and is designed specifically for phone deployment.

Apple’s own on device models reach this threshold (benchmarking at qwen2.5, which seems comparable to gpt4o on lmarena)

https://machinelearning.apple.com/research/introducing-apple-foundation-models

Underpriced:Will there be an AI language model that strongly surpasses ChatGPT and other OpenAI models before the end of 2025

Mistral Small seems similar in performance to GPT 3.5: https://mistral.ai/news/la-plateforme/

Should be a matter of days until someone runs it on their iPhone

It runs on a MacBook, fairly trivial to get ~3x gain in tflops and 3-10x gain in model performance over this window

https://twitter.com/simonw/status/1634635007712165888?s=46&t=7MQWoR0BbBi0qMI9dbdAwQ

Federated learning

Might happen much sooner than I expected

https://twitter.com/harmlessai/status/1626769581858758661?s=46&t=-aOs5vi8y_5tlqgbjNoRKA

As evidence in favour:

• this type of advance in making existing transformers smaller at same performance: https://twitter.com/arankomatsuzaki/status/1624947959644278786?s=20&t=sWoi47Zz-RRprcyDC-jZog

• optimal training data vs parameter count (a la Chinchilla)

• the hefty "neural engines" that Apple is continuously making better in their apple silicon

• that we may find new architectures that work better for language