David J.

Michelle Irvine talks about an important aspect in the daily life of #technicalwriters and #technicalcommunicators - the difficulty in conveying the impact of #content in terms of #metrics that convert to #ROI. Watch this recorded #video from the #WritetheDocs Portland conference 2024 to understand how #documents have an extensive and #measurable impact in your #product deliverables ecosystem.

2

Absolutely fascinating new paper from Anthropic: https://lnkd.in/eq6bQwhw -- this is something that I've been very interested in for a long time -- basically, how do we understand how an LLM represents concepts, and how does mapping and tweaking the weight of those concepts affect downstream behaviours. This points towards new ways in which curated taxonomies, ontologies and entity graphs can point the way to careful concept weighting wihtin an LLM.

6

2 Comments

How do you format your Few-Shot Prompts? We almost exclusively use: Q: <Input> A: <Output> But would love to hear what others use. With Nathan Labenz on https://lnkd.in/egdXwCpy

11

Here are the key points from the paper "Scaling Monosemanticity: Extracting Interpretable Features from Claude 3 Sonnet" by Anthropic: - Anthropic has scaled their sparse autoencoder (SAE) method to extract interpretable features from the large language model Claude 3 #sonnet. This allows them to decompose the model's activations into a sparse set of "feature directions" that represent abstract concepts. - The SAE consists of an encoder that maps model activations to a high-dimensional feature space, and a decoder that reconstructs the original activations from the feature activations. Training minimizes reconstruction error and an L1 penalty to encourage sparsity. - Anthropic found a diversity of highly abstract features in Claude 3 Sonnet, including features for famous people, countries, cities, and code type signatures. Many features are multilingual, and multimodal, and encompass both abstract and concrete instantiations of concepts. - The optimal number of features and training steps for the SAE scales as a power law with the compute budget used to train it. Larger SAEs achieve lower reconstruction losses. - Analyzing the learned features reveals that they become less specific as their activation strength decreases, possibly reflecting the model's confidence in a concept being present. - Extracting interpretable features from large language models like Claude 3 Sonnet could make them safer by revealing how specific features relate to concepts and affect the model's behavior. In summary, Anthropic has made significant progress in scaling its interpretability techniques to #llms, extracting millions of interpretable features that provide insight into the models' inner workings and representations. Source: https://lnkd.in/gqCsGR-Z

4

latest edition explores implications around: (1) making sense of the rapid displacement of startups and “the best model” rat race, (2) how “artifacts of humanity” and Content Credentials will become a human flex of sorts… https://lnkd.in/ecGZKMvr

31

4 Comments

Why graduate programs should be creative in designing discussion board (DB) questions directly incorporating gpt4o as part of designing the DB questions, confronting the usage of GPT4o. Compare the performance of all these gen tools, and why this is important. It is great to see the competition is picking up much faster trying to beat GPT4o

11

3 Comments

Here’s my just-released podcast interview with Jordan Harbinger about my new book and about what the intersecting AI, genetics, and biotechnology revolutions mean for the future of life. #AI #technology #biotechnology

10

1 Comment

Seems important to post here, since this continues to be the largest existential threat to humanity in my estimation: https://lnkd.in/gSKv875R I suppose there are two ways to read the departure of the entire "alignment" / AI safety team: - OpenAI doesn't really care about safety, because achieving AGI is their goal regardless of the consequences - AGI alignment is effectively impossible, and the researchers are just burning cycles for a fundamentally intractable problem I'd guess the latter is accurate, although it sounds like the former is more the cause for the departures.

1

Using ChatGPT as a Libertarian ChatGPT can be used to help strategize the advancement of political principles -- in this case, libertarianism, by researching limitations of a government's capacity to enforce regulations. https://lnkd.in/gVFSUQAE

1

It was great to be part of this balanced discussion about the role of AI in reaching global Net Zero targets, things really aren't as simple as we might first think. Scientists, researchers and industries are making progress in important fields of material science, renewable energy and battery technology to help meet the complex challenges of hitting net zero targets. The field of AI is a genuine contender as a major technology advance to help us forge a path to net zero.

11

Last week, I spoke about AI and regulations at an event at the U.S. Capitol attended by legislative and business leaders. I’m encouraged by the progress the open source community has made fending off regulations that would have stifled innovation. But opponents of open source are continuing to shift their arguments, with the latest worries centering on open source's impact on national security. I hope we’ll all keep protecting open source! I’m encouraged by the progress the U.S. federal government has made getting a realistic grasp of AI’s risks. To be clear, guardrails are needed. But they should be applied to AI applications, not to general-purpose AI technology. Nonetheless, some companies are eager to limit open source, possibly to protect the value of massive investments they’ve made in proprietary models and to deter competitors. It has been fascinating to watch their arguments change over time. For instance, about 12 months ago, the Center For AI Safety’s “Statement on AI Risk” warned AI could cause human extinction and stoked fears of AI taking over. This alarmed leaders in Washington. But many people in AI pointed out that this dystopian science-fiction scenario had little basis in reality. About six months later, when I testified at the U.S. Senate’s AI Insight forum, legislators no longer worried much about an AI takeover. Then the opponents of open source shifted gears. Their leading argument shifted to the risk of AI helping to create bioweapons. Soon afterward, OpenAI and RAND showed that current AI does not significantly increase the ability of malefactors to build bioweapons. This fear of AI-enabled bioweapons has diminished. To be sure, that bad actors could use bioweapons — with or without AI — remains a topic of great international concern. The latest argument for blocking open source AI has shifted to national security. AI is useful for both economic competition and warfare, and open source opponents say the U.S. should make sure its adversaries don’t have access to the latest foundation models. While I don’t want authoritarian governments to use AI, particularly to wage unjust wars, the LLM cat is out of the bag, and authoritarian countries will fill the vacuum if democratic nations limit access. When, some day, a child asks an AI system questions about democracy, the role of a free press, or an independent judiciary, I would like the AI to reflect democratic values rather those of authoritarian leaders. I came away from Washington optimistic about the progress we’ve made defending open source. Looking beyond the U.S. federal government, there are many jurisdictions. Unfortunately, arguments for regulations that would stifle AI continue to proliferate. But talking to regulators does have an impact. If you have a chance to talk to a regulator, I hope you’ll do what you can to help governments better understand AI. [Original text in The Batch (with links): https://lnkd.in/g5SZ6ypa ]

1,613

58 Comments

A backbone of AGI claims is that with enough data, new and untrained capacities for "knowing" emerge within an AI system. Here's a fantastic examination of claims to "emergent properties" within LLMs from Anna Rogers: "We are lacking detailed studies (or even a methodology) to consider the exact relation between the amount and kinds of evidence in the training text data for a particular model output. [Based on the definition that "a property is emergent if the model was not exposed to training data for that property,"] emergent properties are a machine learning equivalent of alchemy – and the bar for postulating that should be quite high. Especially in the face of evidence to the contrary." Dr. Rogers then goes on to examine that contrary evidence in clear and precise language. Worth the read! (And if you prefer more technical papers, this one from 2023 proposes a similar set of challenges to the "emergent properties mirage": ➡ https://lnkd.in/gsqnfeR7)

23

1 Comment

So Kafkaesque! This review by a former KGB illegal compares me to Kafka. Others recently cited John Le Carre and Somerset Maugham and a professor of literature compared me to Dostoevsky - but Amazon rejected his review (below). Like "The Trial," they sent a list of dozens of possible crimes against community standards with no specifics and said, pick one, you are guilty, whatever you pick. (Goodreads published it with nary a hitch.) the review: I operated in the "wilderness of mirrors" as a KGB illegal for ten years during the Cold War. When you do that, you forget who you are and seem to live in an alternate reality--like Mobius. Thieme had to have worked in intelligence to have such an understanding of the moral challenges we faced. Mobius is the Forest Gump of the intelligence community. This fascinating "memoir" is as real as it is surreal and keeps you guessing from start to finish. I watched Thieme's entire Defcon speech on the traumatic impacts of intelligence and security work - Playing Through the Pain: The Dark Side of Professional Intelligence and Security Work - on youtube). I cannot remember having ever heard anything so deep and intrinsically moral, inclusive of the human dilemma. This level of intellect is hard to find across all media these days. The hero of the memoir is a composite of several agents. There are many parallels to my past reality. The deeper I got into the book, the more I was reminded of Kafka. "The Castle" and "The Trial" are two of my favorite books. Alienation is all too real. I also get a touch of nihilism. (Ecclesiastes 1-2). Of course, it is always possible that I am looking in the mirror. We now lie much more often than we speak the truth. And when you do speak the truth people probably think: "What is he hiding?". It gets really confusing when we mold ourselves according to the (perceived) perception of others. Before you know it, we have self-inflicted multiple personality disorder. I can prove that I had at least two permanent personalities who became one when I came out of the closet. My wife never knew which personality would show up. She said I was a chameleon. Thieme's description of PTSD is haunting. I think I understand everything as I read, but then pause and wonder. Do I really? If this were a real memoir, no problem. If I just got it as raw fiction, no problem either. But now that I know the author I can't help guessing how much of himself he put in there. In the end I come away with a ton of mixed thoughts and feelings. It is almost like taking Kafka and rewriting it to make it into a Kafkaesque transcription - Kafka squared!

6

2 Comments

Srijanie Dey, PhD gives a visual deep dive into Anthropic’s Sparse #Autoencoders for LLM #Interpretability in her newest article from Towards Data Science:

3

My new story! Researchers break through our understanding of how neural network-based generative models perform against traditional data sampling techniques in complex systems, unveiling both challenges and opportunities for AI's future in data generation. The study was led by Lenka Zdeborova at EPFL. #AI #neuralnetworks #complexdata #datasampling #datageneration #EPFL

2

BLOGGED: The R&D Library Team: https://lnkd.in/g_tuq-eE The Library team is the team given responsibility for managing reusable artifacts within the company--libraries, tools, frameworks, services (cloud or on-prem), whatever--because achieving actual reusability, it turns out, requires a degree of commitment of time and resources to maintaining and homogenizing those artifacts. Doing this work, at the same time a team is trying to manage a product release, ends up becoming a distraction to the product release cycle, particularly as the shared "thing" becomes more widely shared.

3

This recent study on Whisper's hallucination problems is super interesting (and somewhat concerning): https://lnkd.in/gAqRy95U > "in one sound bite, Whisper correctly transcribed a single, simple sentence, but then hallucinated five additional sentences that contained the words “terror,” “knife” and “killed,” none of which were in the original audio" At Rev, accuracy and reliability is non-negotiable. Unlike other transcription services that might ‘hallucinate’ extra sentences, we ensure every word transcribed is exactly what was spoken. #rev #ai #asr #speechtotext #stt #languagetechnologies #speech #whisper #openai

47

8 Comments