GPT 4 benchmarks, price reduction, rope scaling trick, and AI engineer self-determination.

🎶 Happy birthday to you, happy birthday to you, happy birthday chat GPT-eeeeeeee, happy birthday to you.

Hey everyone, welcome to this special edition of ThursdAI where you're probably gonna have two emails and two podcast episodes today and you can choose which one you want to but we actually recorded both of them live it just they went a little long.

ThursdAI - The only podcast that brings you yearly recaps since chatGPT was released (😂)

This one is the more celebratory one, today is one year from the release of chat GPT and we (and by we I mean I, Alex) decided to celebrate it by recapping not just the last week in AI but the last year (full timeline posted at the bottom of this newsletter)

Going month by month with a swoosh sound in the editing and covering the most important thing that happened in LLM and open source LLMs since chatGPT was released and imagination unlocked the capability for everyone!

We also covered Meta stepping in with Lama and then everything that happened since then in the multi modality and vector databases and agents and everything everything everything, it was a one hell of an hour and a half, we had almost 1K audience members! and so I recommend you listen to this one first and then the week updates later because there were some incredible releases this week as well! (as there are every week)

I think it's important to do like a Spotify wrapped type thing for AI, for something like a one year for chat GPT and I think we'll be doing this every year so hopefully in the year we'll see you here on November 30th covering the next year in AI.

And hopefully the next year in AI system will actually help me summarize all this because it's a lot of work but with that I will just leave you with the timeline and no notes and you should listen to everything because we talked about everything live!

I hope you enjoy this special birthday celebration! (OpenAI sure did, check out this incredibly cute little celebration video they just posted)

Here’s the full timeline with everything important that happened month by month that we’ve covered:

* December 2022 - ChatGPT becomes the fastest growing product in history

* GPT3.5 with 4K context window, instruction finetuning and conversational RLHF

* January

* Microsoft invests additional $10B into OpenAI (Jan 23, Blog)

* February

* LLaMa 1 - Biggest Open Source LLM (February 24 - Blog)

* No commercial license

* 30% MMLU

* No instruction fine-tuninig (RL;HF)

* ChatGPT unofficial APIs exist

* March (the month of LLM superpowers)

* ChatGPT API (March 1, announcement)

* Developers can now build chatGPT powered apps

* All clones so far were completion based and not conversation based

* LLama.cpp from ggerganov + Quantization (March 10, Blog)

* Stanford - Alpaca 7B - Finetune on self-instruct GPT3.5 dataset (March 13, Blog)

* GPT4 release + chatGPT upgrade (March 14 - GPT-4 demo)

* 67.0% HumanEval | 86.4% MMLU

* 8K (and 32K) context windows

* Anthropic announces Claude + Claude instant (March 14 - Blog)

* 56.0% HumanEval

* Folks previously form OAI leave to open Anthropic as research, then pivot from research into commercial

* LMSYS Vicuna 13B - Finetuned based on shareg.pt exports (March 30, Blog)

* April (Embedings & Agents)

* AutoGPT becomes the fastest github starred project + writes it's own code (April 1, Blog)

* Agents start to pop up like mushrooms after the rain

* LLaVa - Multimodality open source begins (April 18, Blog)

* CLIP + Vicuna smushed together to get LLMs eyes

* Bard improvements

* May (Context windows)

* Mosaic MPT-7B with 64K context, 1T parameters, commercial license (May 5, Blog)

* Anthropic updates Claude with 100K context window (May 11, Blog)

* LLongBoi summer begins (Context windows are being stretched)

* Nvidia shows Voyager agents that play Minecraft + Memory stored in Vector DB (May 27, Blog)

* June

* GPT-3.5-turbo + functions API (June 6, Blog)

* GPT3.5 and 4 got a boost in capabilities and steer-ability

* Price reduction on models + 75% reduction on ada embeddings model

* LLaMa context window extended to 8K with RoPE scaling

* AI Engineers self determination essay by swyx

* July

* Code Interpreter GA - ChatGPT can code (July 11, Blog)

* Anthropic Claude 2 - (July 11 - Blog)

* 200K context window

* 71% HumanEval

* LLaMa 2 (July 18 - Blog)

* Base & Chat models (RLHF)

* Commercial license

* 29.9% Human Eval | 68.9% MMLU

* August

* Meta releases Code-LlaMa, code finetune models

* September

* DALL-E 3 - Adds multi-modality on output and chat to image gen (Sep 20, Blog)

* Mistral 7B top performing open source LLM via torrent link (Sep 27, Blog)

* GPT4-V (vision & voice) - Adds multimodality on input (Sep 27, Blog)

* October

* OpenHermes - Mistral 7B finetune that tops the charts from Teknium / Nous Research (Oct 16, Announcement)

* Inflection PI gets connected to the web + supportPi mode (Oct 16, Blog)

* Adept releases multimodal FuYu 8B (Oct 19, blog)

* November

* Grok from Xai - with realtime access to all of X content

* OpenAI dev day

* Combined mode for MMIO (multi modal on input and output)

* GPT-4 Turbo with 128K context, 3x cheaper than GPT-4

* Assistants API with retrieval capabilities

* Share-able GPTs - custom versions of GPT with retrieval, DALL-E, Code Interpreter and vision

* Chatbots with real business use-cases, for example WandBot (that we just launched today! Blog)

* Has vector storage memory

* Available via Discord/Slack

* And custom GPT!

* Microsoft has copilot everywhere in office

Aaaand now we’re here!

What an incredible year, can’t imagine what the next year holds for all of us, but 1 thing is for sure, ThursdAI will be here to keep you all up to date!

P.S - If you scrolled all the way to here, DM me the 🎊 emoji so I know you celebrated with us! It really helps me to know that there is at least a few folks out of the thousands that get this newsletter that scrolls all the way through!

This is a public episode. If you'd like to discuss this with other subscribers or get access to bonus episodes, visit sub.thursdai.news/subscribe