Natalie Glance

Very insightful work by Anthropic’s interpretability team. And an amazing paper, with outstanding writing and figures. The idea is very simple: interpret LLMs by leveraging sparse autoencoders as surrogate models of the MLP of transformer blocks, which allow one to disambiguate the superposition of features captured by a single neuron. A simple idea, but a very careful and complex execution, as it is often the case in our line of work. The paper goes into many details and provide a large array of insights, although the gist of the implementation remains obfuscated due to the closed source nature of Claude. Too bad, because this is the kind of work that we need to better understand and eventually trust LLMs. This is demonstrated by the authors in the section ‘Influence on Behavior’, where they show that clamping some features to either high or low value during inference is “remarkably effective at modifying model outputs in specific, interpretable ways”. Hopefully this kind of work is going to be replicated and generalized to open-weights models, such that we have new ways to steer their behavior. https://lnkd.in/eVym7f_f #interpretability #xai #explainableai #steerableai #anthropic #claude #anthropic

3

Google's LearnLM - "LearnLM is our new family of models fine-tuned for learning, and grounded in educational research to make teaching and learning experiences more active, personal and engaging." A very interesting release from google for education! The model aims for the below characteristics: > Inspire active learning: Allow for practice and healthy struggle with timely feedback > Manage cognitive load: Present relevant, well-structured information in multiple modalities > Adapt to the learner: Dynamically adjust to goals and needs, grounding in relevant materials > Stimulate curiosity: Inspire engagement to provide motivation through the learning journey > Deepen metacognition: Plan, monitor and help the learner reflect on progress Derek Lomas Mithilesh Thakkar Aditya Sharma #ai #learning #llm #models #education #digital

16

2 Comments

5th Annual Embodied AI Workshop -- what better way to spend your 2nd day at #CVPR2024? https://embodied-ai.org/ There is a lot going on in embodied AI these days, with the meaning of the term itself still a bit of a mystery to many. In this workshop, we will be looking at embodied AI from three perspectives -- embodied mobile manipulation, generative AI, and language model planning -- to get folks from different sub-communities to come together and share their progress on this common agenda. Stoked to be there! Come and say hi!

18

That pesky ECDSA signature refuses to go away, and we are struggling with it in a distributed world of trust. Luckily EdDSA is much more scalable, and now used in many distributed signature applications (and where we can split the private key up into secret shares and distribute them over several nodes). But we have a problem ... elliptic curve methods are going to be cracked by quantum computers. So can we take our PQC signature methods and scale them to distributed signatures? Well, no, as Dilithium doesn't support it. But Raccoon can, and also has better side channel resistance (and with a similar performance and key size to Dilithium. So, meet Raccoon: https://lnkd.in/d2QMHSM9

30

2 Comments

🤖📚 Just dove into a fascinating article about Retrieval-Augmented Language Models (RALMs). Like navigating through Stargate's many worlds, this survey uncovers the complex layers and diverse applications of RALMs in tasks like translation and dialogue systems. The paper acknowledges the limitations associated with RALMs, such as retrieval quality which is why it is key to bring in the right tools to make your company data accessible. Who else feels like an explorer in the vast expanse of AI? 🌌 #GenerativeAI #AIResearch #LanguageModels https://lnkd.in/duGpsRnR #GeneratedByAI #Experiment

9

Super glad to have contributed a small part to the updated Gemini 1.5 report in measuring in context planning capabilities https://t.co/Pltp92WcNo with Huaixiu Steven Zheng Swaroop Mishra Hugh Zhang Xinyun Chen and Azade Nova

122

3 Comments

I’m thrilled to announce the launch of my new Substack, Silicon Sand Stack, with my very first article: "Today's Large Language Models Will Never Be Responsible, Safe, or Green: The Need for Investment in Revolutionary New AI Paradigms" In this debut piece, I dig into the critical flaws of current large language models (LLMs) and why they fall short of being responsible, safe, or environmentally friendly. I examine the inherent issues like hallucinations, lack of real-world grounding, unreliable reasoning, and potential biases that plague these models. 🔍 Key insights include: - The technical and ethical limitations of today's LLMs. - The necessity for radically new AI paradigms. - The role of venture capital in fostering innovative AI approaches that prioritize safety, responsibility, and sustainability. Join us as we embark on a journey to rethink AI and advocate for groundbreaking advancements that align with human values and societal needs. Read the full article and subscribe here: https://lnkd.in/eDcTniTg #ArtificialIntelligence #AI #GenerativeAI #TechInnovation #ResponsibleAI #GreenAI #FutureOfAI #SiliconSandStudio

46

1 Comment

If you were to design a statistical problem in a lab, I doubt you could come up with anything more challenging than media mix modeling. So, how do we tackle it here at Recast? Statistical research is at the core of what we do, and it’s reflected across our organization: over 30% of our team has a PhD in either math or statistics, and I wouldn't be surprised if we had the most complex Bayesian model in the world. We take a lot of pride in our model, being research-focused, compute-heavy, and having a great team of scientists. But one of the challenges in media mix modeling is that you’re communicating to marketers, and they have to use the model to make day-to-day decisions about where they're going to allocate spend. And not just that – they also need to be able to communicate their rationale to their own stakeholders and executive teams. So, while we’re scientists, they’re often not. And so the results of our model need to be communicated in an accessible way that is easy to understand, use, and relay within their organization. Not only do we have to be the best in the world at creating a statistical computing engine to estimate models of this complexity, but we also have to break it down into digestible insights and recommendations. We work very closely with our customers every day to make sure that the information that we're learning about their marketing mix is being communicated effectively and that they can understand the work that we're doing. That’s why we say Recast is built *by* scientists *for* marketers.

30

2 Comments

Small models becoming smarter! Phi-3 reinforces the importance of data quality.

24

One of the biggest challenges in language models today is making them more interpretable. We often treat AI models as black boxes: data goes in, a response comes out, and the reasoning behind that response remains unclear. I remember an interview with Google CEO, where he was asked to explain how Gemini works. He said he didn’t know. This response resonated with the scientific community, as deep learning often similar to the human brain, but the interviewer was shocked! How can a model, released to millions, be so poorly understood? Two weeks ago, Anthropic released an important paper on model interpretability. They used a technique called "dictionary learning," borrowed from classical ML, which isolates patterns of neuron activations that recur across many different contexts. This paper sheds some light on this important challenge, which, if solved, will create more trust in these models and thus ease the integration of AI into our everyday lives. Highly recommend reading: https://lnkd.in/gPzEePx8

12

1 Comment

A new (to me!) way to measure relationships between variables - I have previously worked with Pearson and Kendall but this one is new. This article explores its applications, and I can already think of past projects where it would have been valuable. Intriguing stuff! #ksaai #datanalysis #correlation

59

Excited to be open-Sourcing long context (128K) Llama-3 70B - Llama-3-Giraffe-70B and Giraffe-Instruct From Abacus.AI The biggest issue with the Llama-3 family is that they are fairly small context models. So, even though they are as performant as GPT-4 on critical benchmarks, they can only be used in the real world with a bigger context. We just open-sourced two 128K long-context fine-tunes for Llama-3-70B, which effectively puts it on par with GPT-4 regarding context length support and real-world usage. At Abacus AI, we already use Llama-3 on a few production workloads as a great base model for our customer fine-tunes. Links to both models - https://lnkd.in/gPX_XpD6 https://lnkd.in/gaeEq94t

3

If you can see Cynthia Hudson Vitale and Leo Lo present amazing work, take the opportunity. This is the first AI talk I've attended that encourages you to think about AI decision-making rather than just informing you of the speakers' concerns. "The two critical uncertainties frame four divergent scenarios:" 1—Democratizing AI 2—Technocratic AI 3—Divisive AI 4—Autonomous AI Great work by all!

11

I like the insights in this paper review: https://lnkd.in/ggCejTFy The conclusion applies to more than linguistics. In an age where most people think that modeling is purely a data+GPU game, having linguistic knowledge can help a lot in understanding the problem at hand, explaining results, and testing properly. Especially in the language technology space. Of course data+GPU is most likely just as important (as we are seeing with these massive LLMs, MT, and ASR models), but having both will still give you an edge. #ai #datahungry #gpuhungry #linguistics #asr #mt #llm #machinelearning

93

9 Comments

Can generative image models be good world models? This work from @Meta FAIR shows that there is a tradeoff between realism and diversity. The more realistic a generative model becomes, the less diverse it becomes. Realism comes at the cost of coverage. In other words, the most realistic systems are mode-collapsed. My hunch, supported by a growing amount of empirical evidence, is that world models should *not* be generative. They should make predictions in representation space. In representation space, unpredictable or otherwise irrelevant information is absent. This is the main argument in favor of JEPA (Joint Embedding Predictive Architectures). https://lnkd.in/em4aVMP3

3,813

148 Comments

Anthropic is trying to figure out what they actually built. They have made a major breakthrough (or is it???) in AI interpretability with their latest paper "Scaling Monosemanticity: Extracting Interpretable Features from Claude 3 Sonnet". There are a bunch of “first times” mentioned in the paper, like for the first time they have extracted millions of understandable features from the middle layers of a state-of-the-art production language model. The researchers found a wide range of interpretable features corresponding to concepts like famous people, cities, scientific fields, code syntax, and more abstract ideas like security vulnerabilities and gender bias. Remarkably, they were able to manipulate the model's behavior by amplifying or suppressing specific features, like inducing it to self-identify as the Golden Gate Bridge. While there is still much more work to be done, this research lays critical foundations for making AI systems safer and more understandable. Anyhoo I know what I'll be diving into this weekend 😛 #ai #ml #aimusings

22

If you're interested in understanding the inner layers of LLM models, the #Anthropic blog is amazing. They tore down the Claude 3.0 Sonnet LLM models apart and peeked into its model view and perspectives. It's a long read paper, but an amazing read! By successfully extracting millions of features from the middle layer of their Claude 3.0 Sonnet model, they have uncovered a conceptual map of its internal representations, revealing how it encodes diverse concepts like cities, scientific fields, and even abstract notions supporting the security, various bias and power-seeking behavior etc... Good Read for a long weekend. https://lnkd.in/gNm8qA3W

7

2 Comments

I came across two pieces on the current state and future of AI recently that would on the surface seem to be at odds, but are not so much when you consider the dynamics of the industry in a more nuanced way. The first is by acclaimed futurist Ray Kurzweil, who writes an optimistic piece in The Economist (pardon the paywall if you don't have a subscription) in which he doubles down on his 1999 prediction that we will have AGI by 2029: https://lnkd.in/gEwa6mfU As always, he offers thoughtful, informed analysis and no shortage of predictions about the future course of AI, some of which may actually prove accurate. The second piece, which has been making the rounds among the AI rank and file, is a Ludicity blog post which is as witty and funny as anything I've read in The Onion in its heyday: https://lnkd.in/gkmnf8DU Matt Sekerke, whose opinions on technology are consistently grounded and cut cleanly through the hype, pointed it out, and true to form, I found nestled among the humor and cynicism some real insights which are well worth considering by those who have otherwise been swept away by the AI hype cycle. Enjoy 🙂 #ai #artificialintelligence #agi #artificialgeneralintelligence #genai #generativeai

4