Rohan Taori (@rtaori13) / X

Rohan Taori

444 posts

Rohan Taori

@rtaori13

research @AnthropicAI | phd from @StanfordAILab🌲| proud @Cal bear 🐻 | taught w @BerkeleyML

San Francisco

Joined November 2014

Rohan Taori
@rtaori13
Mar 13, 2023
SUPER excited to share our model Alpaca!! ✅ Handles diverse instructions ✅ Simple, reproduceable tuning procedure ✅ Easy to train+host (only 7B params) ✅ Released data pipelines Times are changing. This demo really shows how accessible it is to instruction tune capable LMs.
This post is unavailable.
161K
Rohan Taori
@rtaori13
Dec 8, 2020
Reliability is a key challenge in ML. There are now dozens of robust training methods and datasets - how do they compare? We ran 200+ ImageNet models on 200+ test sets to find out. modestyachts.github.io/imagenet-testb… TDLR: Distribution shift is *really* hard, but common patterns emerge.
Rohan Taori
@rtaori13
Sep 6, 2022
This is a pretty bad take. Stanford doesn’t face the insane budget issues that a public univ like Cal does (been to both schools, experienced the difference firsthand). Govt not prioritizing tax $$ for edu while CS popularity exploding. Stanford also charges 4x the tuition.
Christopher Manning
@chrmanning
Sep 6, 2022
Meanwhile at @Stanford, we just encourage all students to take as many CS courses as they would like …
Rohan Taori
@rtaori13
Sep 2, 2020
Replying to @Uber and @twitter
Lol, looks like no rides in California either way
Rohan Taori
@rtaori13
Mar 15, 2023
🔥🔥 Training code (and data) for Alpaca is now RELEASED! github.com/tatsu-lab/stan… 🔥🔥 Incredibly quick work by @lxuechen @Tianyi_Zh. If you have access to LLaMA, you can now train your own Alpacas!! We also added more capacity to the demo, try it out! crfm.stanford.edu/alpaca/
Rohan Taori
@rtaori13
Mar 13, 2023
SUPER excited to share our model Alpaca!! ✅ Handles diverse instructions ✅ Simple, reproduceable tuning procedure ✅ Easy to train+host (only 7B params) ✅ Released data pipelines Times are changing. This demo really shows how accessible it is to instruction tune capable LMs.
30K
Rohan Taori
@rtaori13
Jul 20, 2021
Any ML system that operates in the real world will inevitably face out-of-distribution (OOD) data that differs from the training set. Inconsistency isn't ideal - so how does OOD performance relate to in-domain performance? We tested a range of models + datasets to find out:
Rohan Taori
@rtaori13
May 23, 2023
2 months ago we brought you Alpaca... today I'm very excited to share the AlpacaFarm!! 🦙🦙🦙 AlpacaFarm is a complete, validated simulator for tuning for instruction-following models: ✅ Cheap (<$200) ✅ FAST (compared to hiring crowdworkers) ✅ Correlates w human data (ρ=0.98)
Tatsunori Hashimoto
@tatsu_hashimoto
May 23, 2023
We are releasing AlpacaFarm, a simulator enabling everyone to run and study the full RLHF pipeline at a fraction of the time (<24h) and cost (<$200) w/ LLM-simulated annotators. Starting w/ Alpaca, we show RLHF gives big 10+% winrate gains vs davinci003 (crfm.stanford.edu/2023/05/22/alp…)
35K
Rohan Taori
@rtaori13
Jul 7, 2023
Here's an idea: release the weights! It's clear these models lag behind in capability, so they don't present a business threat, and the weights would go a long way towards reproducibility of the many experiments the community has run on them 😃 What do you think? @sama @AlecRad
OpenAI
@OpenAI
Jul 6, 2023
GPT-4 API is now available to all paying OpenAI API customers. GPT-3.5 Turbo, DALL·E, and Whisper APIs are also now generally available, and we’re announcing a deprecation plan for some of our older models, which will retire beginning of 2024: openai.com/blog/gpt-4-api…
39K
Rohan Taori
@rtaori13
Feb 27, 2024
Looking for job opps again after 5 years (prev before my phd) and wow, the typical coding interview feels so.... behind the times? Like, what are we testing for by asking someone to implement a sandboxed, python function in 1 hr that GPT4 could do 80% of anyways?
19K
Rohan Taori
@rtaori13
Sep 15, 2022
🎉 The last few weeks have seen the release of #StableDiffusion, #OPT, and other large models. ⚠️ But should we be concerned about an irreversible influx of AI content on the internet? ⚙️ Will this make it harder to collect clean training data for future AI models? 🧵👇 1/6
GIF
Rohan Taori
@rtaori13
Jan 20, 2024
trying to figure out health insurance options while debugging training runs got me like HMO, PPO, DPO, IPO, KTO...
GIF
8.6K
Rohan Taori
@rtaori13
Mar 21, 2024
improving LLM "reasoning" is the new "robustifying" image classifiers
8.5K
Rohan Taori
@rtaori13
Mar 1, 2024
Things in AI move fast, it’s sometimes hard to step back and appreciate the bigger picture. Gemini and Sora make it clear that video is the next frontier 📷 Here's my take on training for video, what it unlocks, and how we get there: rohantaori.com/blog/long-cont…
15K
Rohan Taori
@rtaori13
Apr 18, 2024
Replying to @ethanCaballero
MoEs don't make sense for running models locally (memory is big bottleneck).... I wouldn't do it at 8B scale, 70B scale maybe? but quantization keeps getting better & better
6.3K