Our new goal is to solve alignment of superintelligence within the next 4 years.
OpenAI is committing 20% of its compute to date towards this goal.
Join us in researching how to best spend this compute to solve the problem!
With the InstructGPT paper we found that our models generalized to follow instructions in non-English even though we almost exclusively trained on English.
We still don't know why.
I wish someone would figure this out.
Super excited about our new research direction for aligning smarter-than-human AI:
We finetune large models to generalize from weak supervisionโusing small models instead of humans as weak supervisors.
Check out our new paper:
Today we're releasing a tool we've been using internally to analyze transformer internals - the Transformer Debugger!
It combines both automated interpretability and sparse autoencoders, and it allows rapid exploration of models without writing code.
The names for "precision" and "recall" seem so unintuitive to me, I have probably opened the Wikipedia article for them dozens of times.
Does anyone know a good mnemonic for them?
Really exciting new work on automated interpretability:
We ask GPT-4 to explain firing patterns for individual neurons in LLMs and score those explanations.
Before we scramble to deeply integrate LLMs everywhere in the economy, can we pause and think whether it is wise to do so?
This is quite immature technology and we don't understand how it works.
If we're not careful we're setting ourselves up for a lot of correlated failures.
I'm very excited that today OpenAI adopts its new preparedness framework!
This framework spells out our strategy for measuring and forecasting risks, and our commitments to stop deployment and development if safety mitigations are ever lagging behind.
Extremely exciting alignment research milestone:
Using reinforcement learning from human feedback, we've trained GPT-3 to be much better at following human intentions.
Reinforcement learning from human feedback won't scale.
It fundamentally assumes that humans can evaluate what the AI system is doing.
This will not be true once AI becomes smarter than humans.
This is one of the craziest plots I have ever seen.
World GDP follows a power law that holds over many orders of magnitude and extrapolates to infinity (!) by 2047.
Clearly this trend can't continue forever. But whatever happens, the next 25 years are going to be pretty nuts.
This is the most important plot of alignment lore:
Whenever you optimize a proxy, you make progress on your true objective for a while.
At some point you start overoptimizing and do worse on your true objective (hard to know when).
This applies to all proxy measures ever.
We're distributing $1e7 in grants for research on making superhuman models safer and more aligned.
If you've always wanted to work on this, now is your time!
Apply by Feb 18:
An important test for humanity will be whether we can collectively decide not to open source LLMs that can reliably survive and spread on their own.
Once spreading, LLMs will get up to all kinds of crime, it'll be hard to catch all copies, and we'll fight over who's responsible
We're hiring research engineers for alignment work at
@OpenAI
!
If you're excited about finetuning gpt3-sized language models to be better at following human intentions, then this is for you!
Apply here:
Jailbreaking LLMs through input images might end up being a nasty problem.
It's likely much harder to defend against than text jailbreaks because it's a continuous space.
Despite a decade of research we don't know how to make vision models adversarially robust.
Really interesting result on using LLMs to do math:
Supervising every step works better than only checking the answer.
Some thoughts how this matters for alignment ๐
GPT-4 is safer and more aligned than any other OpenAI has deployed before.
Yet it's not perfect. There is still a lot to do to improve safety and we're planning to make updates over the coming months.
Huge congrats to the team on all the progress! ๐
It's been heartening to see so many more people lately starting to take existential risk from AI seriously and speaking up about it.
It's a first step towards solving the problem.
Today was my last day at
@DeepMind
. It's been an amazing journey; I've learned so many things and got to work with so many amazing people!
Excited for what comes next!
Super exciting new research milestone on alignment:
We trained language models to assist human feedback!
Our models help humans find 50% more flaws in summaries than they would have found unassisted.
If you're into practical alignment, consider applying to
@lilianweng
's team. They're building some really exciting stuff:
- Automatically extract intent from a fine-tuning dataset
- Make models robust to jailbreaks
- Detect & mitigate harmful use
- ...
The superalignment fast grants are now decided!
We got a *ton* of really strong applications, so unfortunately we had to say no to many we're very excited about.
There is still so much good research waiting to be funded.
Congrats to all recipients!
Great conversation with
@robertwiblin
on how alignment is one of the most interesting ML problems, what the Superalignment Team is working on, what roles we're hiring for, what's needed to reach an awesome future, and much more
๐ Check it out ๐
The agent alignment problem may be one of the biggest obstacles for using ML to improve peopleโs lives.
Today Iโm very excited to share a research direction for how weโll aim to solve alignment at
@DeepMindAI
.
Blog post:
Paper:
How do we uncover failures in ML models that occur too rarely during testing? How do we prove their absence?
Very excited about the work by
@DeepMindAI
โs Robust & Verified AI team that sheds light on these questions! Check out their blog post:
RSA was published 45 years ago and yet the universally accepted way to digitally sign a document is to make an indecipherable squiggle on a touch screen that no one ever checks.
Everyone has a right to know whether they are interacting with a human or AI.
Language models like ChatGPT are good at posing as humans.
So we trained a classifier to distinguish between AI-written and human-written text.
But it's not fully reliable.
Y'all should stop using logprob-based evals for language models.
I.e. don't craft two reference responses and calculate logP(good response | prompt) - logP(bad response | prompt).
This wouldn't actually measure what you care about!
Very excited to deliver the
#icml2019
tutorial on
#safeml
tomorrow together with
@csilviavr
!
Be prepared for fairness, human-in-the-loop RL, and a general overview of the field.
And lots of memes!
Some statistics on the superalignment fast grants:
We funded 50 out of ~2,700 applications, awarding a total of $9,895,000.
Median grant size: $150k
Average grant size: $198k
Smallest grant size: $50k
Largest grant size: $500k
Grantees:
Universities: $5.7m (22)
Graduateโฆ
One of my favorite parts of the GPT-4 release is that we asked an external auditor to check if the model is dangerous.
This project lead by
@BethMayBarnes
tested if GPT-4 could autonomously survive and spread. (The answer is no.)
More details here:
It supports both neurons and attention heads.
You can intervene on the forward pass by ablating individual neurons and see what changes.
In short, it's a quick and easy way to discover circuits manually.
This problem is sometimes called *scalable oversight*. There are several ideas how to do this, and how to measure that we're making progress.
The path I'm very excited for is using models like ChatGPT to assist humans at evaluating other AI systems.
I'm super excited to be co-leading the team together with
@ilyasut
.
Most of our previous alignment team has joined the new superalignment team, and we're welcoming many new people from OpenAI and externally.
I feel very lucky to get to work with so many super talented people!
I recommend talking to the model to explore what it can help you best with. Try out how it works for your use case and probe it adversarially. Think of edge cases.
Don't rush to hook it up to important infrastructure before you're familiar with how it behaves for your use case.
Submtting a NeurIPS paper and unsure how to write your broader impact statement?
This blog post will guide you through it!
Comes with a few concrete examples, too.
By Carolyn Ashurst,
@Manderljung
,
@carinaprunkl
,
@yaringal
, and Allan Dafoe.
We find that large models generally do better than their weak supervisor (a smaller model), but not by much.
This suggests reward models won't be much better than their human supervisors.
In other words: RLHF won't scale.
There are a lot of exciting things in the Codex paper, but my favorite titbit is the misalignment evaluations by
@BethMayBarnes
: Subtly buggy code in the context makes the model more likely to write buggy code, and this discrepancy gets larger as the models get bigger!
20% of compute is not a small amount and I'm very impressed that OpenAI is willing to allocate resources at this scale.
It's the largest investment in alignment ever made, and it's probably more than humanity has spent on alignment research in total so far.
For lots of important tasks we don't have ground truth supervision:
Is this statement true?
Is this code buggy?
We want to elicit the strong model's capabilities on these tasks without access to ground truth.
This is pretty central to aligning superhuman models.
@ESYudkowsky
We'll stare at the empirical data as it's coming in:
1. We can measure progress locally on various parts of our research roadmap (e.g. for scalable oversight)
2. We can see how well alignment of GPT-5 will go
3. We'll monitor closely how quickly the tech develops
LLMs can hallucinate and lie. They can be jailbroken by weird suffixes. They memorize training data and exhibit biases.
๐ง We shed light on all of these phenomena with a new approach to AI transparency. ๐งต
Website:
Paper:
@michhuan
@OpenAI
@NPCollapse
@ilyasut
Alignment is not binary and there is a big difference between aligning human level systems and aligning superintelligence.
Making roughly human-level AI aligned enough to solve alignment is much easier than solving alignment once and for all
But even our simple technique can significantly improve weak-to-strong generalization.
This is great news: we can make measurable progress on this problem today!
I believe more progress in this direction will help us align superhuman models.
New paper on teaching RL agents to understand the meaning of instructions. Instead of manually specifying rewards, we learn them from goal-state examples. With
@DBahdanau
, Felix Hill,
@edwardfhughes
,
@pushmeet
, and
@egrefen
!
If you are worried about risks from frontier model capabilities, consider applying to the new Preparedness team!
If we can measure exactly how dangerous models are, the conversation around this will become more grounded. Exciting that this new team is taking on the challenge!
We are building a new Preparedness team to evaluate, forecast, and protect against the risks of highly-capable AIโfrom today's models to AGI.
Goal: a quantitative, evidence-based methodology, beyond what is accepted as possible:
What should we be aligning to when we're building AI systems like ChatGPT?
I'm excited about this idea based on simulated deliberative democracy.
Would love to hear what y'all think :)
Interested in getting into machine learning research and AI safety in particular?
@80000Hours
recently interviewed me about this. Check out the podcast:
Incentives are the most powerful force in the universe.
Stronger that any other physical force.
E.g. if you commit enough money to have a train float in the air, it will float.
Had a great time chatting with
@dfrsrchtwts
about our Superalignment plans.
If you want to learn more about what our team is up to and hear my latest thoughts about alignment, check it out:
How do you measure the distance between two reward functions?
Our EPIC distance is invariant to reward shaping, can be approximated efficiently, and is predictive of policy training success and transfer!
New paper with
@ARGleave
,
@MichaelD1729
et al.
What could a once-and-for-all solution to the alignment problem actually look like?
It'll be very different from what we do today.
This is my attempt to sketch it out:
Alignment is fundamentally a machine learning problem, and we need the world's best ML talent to solve it.
We're looking for engineers, researchers, and research managers. If this could be you, please apply:
We're distributing $1e7 in grants for research on making superhuman models safer and more aligned.
If you've always wanted to work on this, now is your time!
Apply by Feb 18:
For comparison: We spent <2% of the pretraining compute on fine-tuning and collect a few 10,000s of human labels and demos. Our 1.3b parameter models (GPT-2 sized!) are preferred over a prompted 175b parameter GPT-3.
I really love this new paper showing how single neurons respond across modalities in CLIP models. Opens up a new avenue of new typographic attacks to fool these kinds of models.
By
@gabeeegoooh
,
@ch402
, and others.
@michael_nielsen
I agree that these questions are important, but we don't need a definitive answer in order to make progress on alignment.
Right now we don't even know how to make them reliably follow anyone's intent, or do things we all agree on.
Mitigating misuse of AI is a different problem.
So many alignment plans revolve around "we'll convince everyone to not do X."
Maybe you can buy some time, but people will do X anyway. We should instead spend our time trying to figure out how to make X aligned & safe.
Multiparty computation is awesome because it lets multiple parties train a model without seeing the weights.
But there are fundamental limits to making it scalable: >24x overhead!
Our new paper addresses this problem.
w/
@MiljanMartic
@iamtrask
et al.
Why 4 years? It's a very ambitious goal, and we might not succeed. But I'm optimistic that it can be done.
There is a lot of uncertainty how much time we'll have, but the technology might develop very quickly over the next few years.
I'd rather have alignment be solved too soon
I'm very interested in techniques for supervising models to do tasks that are difficult for humans to evaluate.
To study this, we trained a model on summarizing entire books!
Read more ๐
I interviewed OpenAI's Head of Alignment
@janleike
on their new superalignment project on which they're spending $100m's in an attempt to figure out how to make superhuman AI not go rogue and wreck everything:
Constitutional AI doesn't let you avoid labeling data by writing down some rules.
You still need to figure out how good your rules are. So you need to label a validation set.
Then you'll get some accuracy on the validation set. How can you increase this accuracy?
It's still early days, but it's been cool to see some interesting trends:
1. Later layers are harder to explain than earlier ones
2. Simple interventions into pretraining can improve explainability of neurons
3. Simple tricks like iterative refinements can improve explanations