Matthias Keller

There are many LLM security papers out there that do something odd and get potentially harmful content out. But rarely is there any clue if this is against policy - ie., is unintended. Reviewers are left to judge in the dark what policy a given LLM was built on. It's a mess. Values and norms come from everywhere and vary everywhere, and without explicit policy known, it's impossible to judge whether that model's policy has been circumvented, or the difficulty/significance of doing so.  For example, if an LLM has no policy, all outputs are intended. It might be unsafe and a generally poor move to build and distribute it, but being able to get bad things out of a model with no policy is NOT a security problem.

97

13 Comments

Now you can give ChatGPT direct access to your Google and Microsoft cloud accounts so it can read your files. Here is a video of Sam Altman at the press conference announcing the feature. #Friday #ActuallyAFeature #DudeAreYouBeingSarcasticIDontEvenKnowAnymore

14

6 Comments

AI at Meta has released the most powerful open-source model yet as of today: 𝑳𝑳𝑨𝑴𝑨3 - 8B and 70B models at 8k context length! They highlight improvements in each of the following aspects as the key differentiator: a) Model Architecture b) Pretraining data c) scaling up pre-training d) Instruction fine-tuning 💥 𝐏𝐫𝐞𝐭𝐫𝐚𝐢𝐧𝐢𝐧𝐠 & 𝐒𝐜𝐚𝐥𝐢𝐧𝐠 𝐮𝐩: ✈️ Vocab of 128k tokens & GQA across both 8B and 70B, 8k native context length ✈️ Pretrained over 15T tokens, 4 times more code than LLAMA2 ✈️ Multi-lingual : over 5% of pretraining data consists of high-quality non English data that covers over 30 languages: don't expect same performance as English ✈️ LLAMA2 generates training data (like in Self-Instruct) to generate text classifiers that filter data powering LLAMA3 ✈️ Both 8B and 70B improve even after the model is trained on 2 orders of magnitude more data i.e log-linear improvement after 15T tokens training ✈️ Training runs on 2 custom-built 24k GPU clusters with an overall training time efficiency of 95% ✈️ 𝘖𝘷𝘦𝘳𝘢𝘭𝘭, 3 𝘵𝘪𝘮𝘦𝘴 𝘮𝘰𝘳𝘦 𝘦𝘧𝘧𝘪𝘤𝘪𝘦𝘯𝘵 𝘵𝘩𝘢𝘯 𝘓𝘓𝘈𝘔𝘈2 💥 𝑰𝒏𝒔𝒕𝒓𝒖𝒄𝒕𝒊𝒐𝒏 𝑭𝒊𝒏𝒆𝒕𝒖𝒏𝒊𝒏𝒈: ✈️ Combination of SFT + PPO + DPO + Rejection Sampling ✈️ 𝑸𝒖𝒂𝒍𝒊𝒕𝒚 𝒐𝒇 𝒑𝒓𝒐𝒎𝒑𝒕𝒔 𝒖𝒔𝒆𝒅 𝒊𝒏 𝑺𝑭𝑻, 𝒑𝒓𝒆𝒇𝒆𝒓𝒆𝒏𝒄𝒆 𝒓𝒂𝒏𝒌𝒊𝒏𝒈𝒔 𝒖𝒔𝒆𝒅 𝒊𝒏 𝑫𝑷𝑶, 𝑷𝑷𝑶 𝒉𝒂𝒔 𝒂𝒏 𝒐𝒖𝒕𝒔𝒊𝒛𝒆𝒅 𝒊𝒎𝒑𝒂𝒄𝒕 𝒐𝒏 𝒎𝒐𝒅𝒆𝒍 𝒑𝒆𝒓𝒇𝒐𝒓𝒎𝒂𝒏𝒄𝒆 ✈️ Careful data curation and multiple rounds of QA on annotations from human annotators ✈️Training on preference ranking improved the model performance greatly on coding and reasoning tasks 💥 𝑳𝑳𝑨𝑴𝑨3 𝑮𝒂𝒖𝒓𝒅𝒓𝒂𝒊𝒍𝒔: ✈️ Updated and new safety tools : LLAMA Gaurd2 & cybersecEval 2 ✈️ CodeShield - inference time gaurdrail for filtering insecure code ✈️ Instruction Fine-tuned model redteamed for safety : By generating adverserial prompts that try to elicit problematic responses. ✈️ LLAMAGaurd - foundational for prompt and response safety, can be easily finetuned to create a new taxonomy 💥 𝑰𝒏𝒇𝒆𝒓𝒆𝒏𝒄𝒆 & 𝑭𝒖𝒕𝒖𝒓𝒆 ✈️ Despite LLAMA3 having 1B more params than LLAMA2 7B, the improved tokenizer efficiency and GQA contribute to maintaining the inference efficiency on par with Llama 2 7B. ✈️ 400B+ parameters, Multilingual, Multimodal, and longer context window LLAMA3 would be available on AWS soon. Try it out! #generativeAI #llm #llama3 #aws #bedrock #sota

33

1 Comment

https://lnkd.in/dJW9sZpx SelfExtend introduces a novel approach to expanding the context window of LLMs. It leverages a two-tiered attention mechanism—grouped attention for long-range dependencies and neighbor attention for local interactions—both derived from the model's original self-attention. This method allows seamless context window extension without requiring fine-tuning, and extensive experiments have demonstrated its effectiveness across various benchmarks.

12

1 Comment

https://lnkd.in/eSrt4ey7 "Llama or Large Language Model Meta AI is a family of LLMs introduced by Meta AI in February 2023. The first version of the model was released in four sizes — 7B, 13B, 33B, and 65 billion parameters. Reportedly, the 13B model of Llama outperformed OpenAI’s GPT-3 which had 135 billion parameters. To simplify, parameters here are a measure of the size and complexity of an AI model and generally, a larger number of parameters means an AI model is more complex and powerful." Meta AI at Meta #Llama #AI #LLM #LargeLanguageModel

1

Prompt engineering has emerged as a critical skill for effectively interacting with large language models. This comprehensive study examines the diverse landscape of prompting techniques, security vulnerabilities, and key considerations for prompt design. The authors have done a great job creating a taxonomical structure of prompt techniques, making it easier for AI practitioners to systematically implement and understand the relationships among them. A highlight for me was the detailed case study on using prompt engineering to detect signals of suicidal crisis in text, showcasing a step-by-step process to improve prompts and enhance results. The study also emphasizes how sensitive LLM responses are to different prompts and the black-box nature of these models, underscoring the need for automated prompting techniques to minimize result variability. Paper - https://lnkd.in/gjHtXmDD #LLM #AI #PromptEngineering

29

Thanks for sharing Baris Yuksel this great site on machine learning trends. https://lnkd.in/d3DTuRBp Next generation AI Systems will need to innovate across different domains at once: Systems - How do we build and run these large systems (cost to train, cost to run) Algorithms - How do we change the architecture such that we can get much better scaling laws.

9

Sebastian Raschka's blog provides invaluable insights and comprehensive information on #LLMs and #LLMOps, making it a valuable resource for anyone seeking knowledge and understanding in this field. #NLP #LLMS #LLMOps

9

✎ Transparent border: https://lnkd.in/epMKzciT Quick rebound off of Dave Rupert's post about the utility of `border-color: transparent`. It's a helpful technique for reducing janky behavior and creating sturdy themes in a design system.

65

7 Comments

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

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

With the announcement of S3-native-streams (Freight clusters), here is a commentary on Confluent strategy regarding object storage, streaming and an open data architecture. https://lnkd.in/dTg6DKVy

48

6 Comments

“You know, you can trust things that you can articulate” As part of our efforts on Trusted AI-assisted programming, the paper demonstrates effectiveness of LLM in generating formal postconditions, can be executed or analyzed with several use cases to be presented in FSE’24 in Brazil with Madeline Endres, Sarah Fakhoury, Shuvendu Lahiri. Our ambition is to equip the AI assisted programming system enough capabilities to fomally analyze (and at the extreme verify) the AI generated code, so that the programmer does not have to blindly trust the system. First step to verification is formalizing the specifications. This paper shows that the specifications (e.g., postcondition of a method) can be effective formalized with AI agents. In future work, we also explore the proof part of the verification with assistance from LLMs. No debate, LLMs do show tremendous emergent capabilities. Modern LLMs are able to answers questions (such as writing a syntactically correct code in a completely made up language, only with few examples), that are impossible to imagine 5/6 years ago. We aim at harnessing such capabilities, while ensuring that we get exactly what we want by imposing constraints, such as specifications, verification tools, etc.

45

Fortune 500 companies using #genai systems like #openai or #github #copilot should go back in history and read up on the Madoff scheme. What allowed the Ponzi to continue? NO 3rd PARTY AUDIT. If you are using #saas or #genai you need a 3rd party audit of what data they are using, or...you might get burned in a similar way....

1

Great to see Anthropic being thoughtful on how to make AI and LLMs responsible with specific framework, investments and teams working on it - this is a bit similar to cybersecurity, disaster recovery, business continuity and security as job 0 in the cloud. Good read.

8

Every Way To Get Structured Output From LLMs “This post will be interesting to you if: -you're trying to get structured output from an LLM, -you've tried response_format: "json" and function calling and been disappointed by the results, -you're tired of stacking regex on regex on regex to extract JSON from an LLM, -you're trying to figure out what your options are. Everyone using LLMs in production runs into this problem sooner or later: what we really want is a magical black box that returns JSON in exactly the format we want. Unfortunately, LLMs return English, not JSON, and it turns out that converting English to JSON is kinda hard. Here's every framework we could find that solves this problem, and how they compare.” https://lnkd.in/dXvjz8tS

8

Post Open. I appreciate Bruce Perens not trying to re-define Open Source, but I'd prefer some other terminology which does not imply replacement. I think it is fantastic to experiment with different kinds of Proprietary licenses, yet if you are to polish Open Source definition it should be around respecting and enhancing freedom of software users, continuing to empower permissionless innovation, not ensure Devs can be paid what they think they deserve to be paid - Proprietary Software licenses already solve this problem well https://postopen.org/ #opensource #postopen

10

1 Comment

Cool article on how Fugaku, Japan’s fastest supercomputer, has gone virtual on AWS. “For these scientific outcomes to proliferate and create an actual impact on society, just building this 160,000-node machine is far from sufficient. The value of Fugaku is not only the hardware, but also the software solutions including the applications. We’ve injected this supercomputing environment directly into the cloud to extend applications and have much larger impact than in a single machine.” #Supercomputing #AWS #cloud Link to the article in the first comment.

23

3 Comments