Using data to craft the best whisky in the world
Joined January 2017
- Tweets4.3K
- Following788
- Followers399
- Likes6.4K
Matthew C. Higgs retweeted
This is the best paper written so far about the impact of AI on scientific discovery
Readers added context they thought people might want to know
There seem to be grave concerns about the integrity of the research.
MIT announced that they "conducted an internal, confidential review and concluded that the paper should be withdrawn from public discourse."
economics.mit.edu/news/assuring-…
wsj.com/tech/ai/mit-sa…
Matthew C. Higgs retweeted
(1/6) Recommender systems shape our digital experiences, filtering much of what we see online. But how should we measure their influence? We provide a unified causal framework to think through this question and develop metrics to audit recommender systems.
arxiv.org/abs/2409.13210
Matthew C. Higgs retweeted
Time to enjoy yet another way GenAI is more of a headache than enterprises realized:
Data deletion takes on a whole new meaning when it’s possible to recover data a company didn’t know it had.
I explain in the blog 👇
bit.ly/quaesita_linkedinopti…
Matthew C. Higgs retweeted
The AI Risk Repository can be found here: airisk.mit.edu/
Matthew C. Higgs retweeted
.@SimonPrinceAI finished 68 (!!!) notebooks that go with his book "Understanding Deep Learning".
These are *excellent* for learning/teaching deep learning.
Notebooks and book pdf at: udlbook.com
#100DaysOfMLCode
Matthew C. Higgs retweeted
This is not good, "Surprisingly, we observe a significant decline in LLMs’ reasoning abilities under format restrictions."
Link: arxiv.org/abs/2408.02442
Matthew C. Higgs retweeted
Building a platform for generative AI applications
huyenchip.com/2024/07/25/gen…
After studying how companies deploy generative AI applications, I noticed many similarities in their platforms. This post outlines these common components, what they do, and implementation considerations.
This post starts from the simplest architecture and progressively adds more components.
1. Enhance context input into a model by giving the model access to external data sources and tools for information gathering.
2. Put in guardrails to protect your system and your users.
3. Add model router and gateway to support complex pipelines and add more security.
4. Optimize for latency and costs with cache.
5. Add complex logic and write actions to maximize your system’s capabilities.
I try my best to keep the architecture general, but certain applications might deviate. As always, feedback is appreciated!
Matthew C. Higgs retweeted
Breakthrough paper, showing that optimal learners in stochastic optimization memorize a constant fraction of their training data.
Mahdi will be talking on Thursday at ICML, poster immediately after.
I am at #ICML2024 presenting our work on the information complexity of stochastic convex optimization with my amazing collaborators.
Oral talk: Thu 25 Jul 11:15 a.m.
Poster: Thu 25 Jul 11:30 a.m.
Arxiv Link: arxiv.org/abs/2402.09327
ICML link: icml.cc/virtual/2024/poster/…
Matthew C. Higgs retweeted
In talking to policy makers and AI researchers, I realised there's a fact agreed upon by all researchers, but understood by almost no policy makers.
This uncomfortable fact is why AI policy is hard.
Matthew C. Higgs retweeted
I still can't get over how bad AI has become. And this is coming from a computer scientist who spent over a decade studying and programming AI.
I fucking LOVE AI, but here are 10 reasons I absolutely fucking hate it now, a 🧵 ...
Matthew C. Higgs retweeted
we removed the AI-powered search from our docs because it was unreliable and made up fake features
still waiting for the production-grade llm that can reliably produce data without hallucinations
long way to go
Matthew C. Higgs retweeted
Whether LLMs can reliably be used for decision making and benefit society depends on whether they can reliably represent uncertainty over the correctness of their outputs. There's anything but consensus. In new work we find LLMs must be taught to know what they don't know. 1/6
Predictions without reliable confidence are not actionable and potentially dangerous. In new work, we deeply investigate uncertainty calibration of large language models. We find LLMs must be taught to know what they don’t know:
arxiv.org/abs/2406.08391
w/ @psiyumm et al.
1/8
Matthew C. Higgs retweeted
This chart is a great rebuttal of the view that RAG will solve hallucination. In the legal domain, it turns out that it does improve precision somewhat, but at the expense of recall. Really great paper by Stanford & Yale researchers assessing legal AI. dho.stanford.edu/wp-content/…
ALT The image is a bar chart comparing the proportion of hallucinated and incomplete answers across different legal research tools. Hallucination: Lexis: ~0.2 Westlaw: ~0.3 Practical Law: ~0.3 GPT-4: ~0.4 Incomplete: Lexis: ~0.25 Westlaw: ~0.3 Practical Law: ~0.6 GPT-4: ~0.1 Caption: Comparison of hallucinated and incomplete answers across generative legal research tools. Hallucinated responses are those that include false statements or falsely assert a source supports a statement. Incomplete responses are those that fail to either address the user’s query or provide proper citations for factual claims.
Matthew C. Higgs retweeted
In a wild follow up to this, Westlaw tried to say the researchers that found 1 in 6 (!) false responses from their AI tools were "auditing the wrong product". So those researchers asked for access the "right" product and found hallucinations in... 1 in 3 responses 🤦🏾♀️
A similar situation happened recently - LexisNexis & Thomson Reuters (parent co of Westlaw) have both released AI products to be used by actual lawyers. Stanford researchers (incl. Dan Ho, @chrmanning ) found that the tech was disastrously inappropriate: hai.stanford.edu/news/ai-tri…
Matthew C. Higgs retweeted
Surprise, surprise, studies show that LLMs don't reduce doctor "burnout" and pose serious harms to patients.
Hardly a shocker, but always good to see the rubber of serious research hit the road of grandiose hype. msn.com/en-us/health/other/a…
Matthew C. Higgs retweeted
IMO @AnthropicAI is very close to making a breakthrough in productizable interpretability.
For ~4 years all we've had to really control LLMs is temperature/top_p and logit bias. We recently got `seed` and constrained structured output, with `interactive=false` on the way.
But now Claude Sonnet can let people clamp up/down 34 million features ranging from famous people and noun categories to abstract concepts like its ability to keep a secret, code errors (!) and its representation of self. This is "representation engineering writ large", and probably easy to retrain the SAEs for any set of features you care about.
Do not sleep on their paper. Ground breaking series of work that led up to today.
By end of 2024, steering foundation models in latent/activation space will outperform steering in token space ("prompt eng") in several large production deployments.
I felt skeptical about this in summer '23, felt vaguely positive in Jan, and now think it's more likely than not, and I'm more optimistic than ever about the direction of my work since 2022 around a joint exploration of interfaces and latent visualization/steering of foundation models. Anthro's published work today is a milestone in a steady march toward this future that started in mid 2023.
Anthropic and DM's leadership in this area, combined with lots of community efforts in work like steering vectors, better sparse autoencoders, various image editing UI prototypes all push us toward the future here, but once technical foundations are there, interface will be much more obviously a bottleneck to utility, alignment, and capability.
I'm very excited about the interface possibilities this will open up, particularly for multimodal models and creative use cases. For a moment I thought it was possible that dialogue may eat everything. I don't think so anymore. We'll see new universes of possibilities in both.
(And if frontier labs don't have serious interface research bets today, this would probably be a good time to reconsider it :-)
Matthew C. Higgs retweeted
If you missed "Two Types of AI Existential Risk" by @Dr_Atoosa it's essential reading.
Building on a complex systems approach it explores how AI could make our societies more fragile to a shock, which would then have long-lasting implications.
arxiv.org/pdf/2401.07836
Matthew C. Higgs retweeted
Really enjoyed LinkedIn's report on what worked and what didn't when deploying LLM applications. 4 takeaways.
1. Structured outputs
They chose YAML over JSON as the output format because YAML uses less tokens. Initially, only 90% of the outputs are correctly formatted YAML. They used re-prompting (asking the model to fix its YAML responses), which increased the number of API calls significantly.
They then analyzed the common formatting errors, added those hints to the original prompt, and wrote an error fixing script. This reduced their errors to 0.01%.
