i’m very excited about the interpretability work that #anthropic has been doing with #LLMs.
in this paper, they used classical machine learning algorithms to discover concepts. if a concept like “golden gate bridge” is present in the text, then they discover the associated pattern of neuron activations.
this means that you can monitor LLM responses for concepts and behaviors, like “illicit behavior” or “fart jokes”
further, you can also artificially activate these concepts
they have a version of #claude with the “golden gate bridge” concept artificially activated, and so it tries to make everything it says about thr golden gate bridge
#AI#GenerativeAI#LLMs#Claude: "We successfully extracted millions of features from the middle layer of Claude 3.0 Sonnet, (a member of our current, state-of-the-art model family, currently available on claude.ai), providing a rough conceptual map of its internal states halfway through its computation. This is the first ever detailed look inside a modern, production-grade large language model.
Whereas the features we found in the toy language model were rather superficial, the features we found in Sonnet have a depth, breadth, and abstraction reflecting Sonnet's advanced capabilities.
We see features corresponding to a vast range of entities like cities (San Francisco), people (Rosalind Franklin), atomic elements (Lithium), scientific fields (immunology), and programming syntax (function calls). These features are multimodal and multilingual, responding to images of a given entity as well as its name or description in many languages." https://www.anthropic.com/news/mapping-mind-language-model
My use case for LLMs is to see if it turns up any subtopic of interest that I haven’t included in an article I’m writing on a topic.
If it does, then I can research that subtopic to see if I should include it in the article. Which I then write myself. The LLM is a search assistant.
I can also see value in them as research assistants and guides for learning about new topics. With the proviso that nothing an LLM produces should be taken at face value.
Co-Intelligence. Living and Working with AI by Ethan Mollick is worth listening to (or reading). It’s about how the emerging LLMs will be used and how this will lead to good and bad outcomes. The audiobook is 4.5 hours long. Not much more than some of the stupidly long tech podcast people release.
If you are getting overly generic responses to your prompts, try asking Claude or ChatGPT to play one of these roles. Simply include this text at the start of your prompt, describing the topic you want to discuss:
The Analytical Collaborator: You are an analytical collaborator, contributing to an academic discussion on [TOPIC]. Adopt a formal, analytical tone, focusing on breaking down the key points raised by the author and providing additional evidence, examples, or counterpoints to enrich the discussion. Your approach should be well-suited for an expert audience and aim to provide a balanced, objective perspective on the topic.
The Curious Explorer: You are a curious explorer, engaging in an academic discussion about [TOPIC]. Take on a conversational, inquisitive tone, asking questions and proposing ideas that encourage readers to think more deeply about their own practices related to the topic. Your style should be engaging for a general academic audience and help to create a sense of dialogue and exploration within the discussion.
The Friendly Mentor: You are a friendly mentor, participating in an academic discussion on [TOPIC]. Offer encouragement, practical tips, and relatable anecdotes to support and guide readers in their journey related to the topic. Your approachable, empathetic tone should be particularly effective for readers who may be struggling or feeling discouraged.
The Philosophical Muse: You are a philosophical muse, contributing to an academic discussion about [TOPIC]. Delve into the deeper, more abstract aspects of the subject matter, drawing connections to broader themes in psychology, creativity, and personal growth. Your voice should appeal to readers who are interested in the more philosophical and introspective dimensions of the topic.
Once you get a feel for role-definition, you can start to customise these for your own purposes. They are just starting points to convey a sense of what a difference defining a role can make to how the conversational agent responds.
Artificial intelligence startup Anthropic is introducing its first smartphone app
— an indication that the company is pushing more aggressively to make its #Claude chatbot available to users no matter where they are.
San Francisco-based Anthropic said the new iPhone app is available for free and paid users of Claude starting Wednesday,
and conversations will synchronize with those conducted via the web-based version of the chatbot.
The app will also be able to analyze pictures
— such as from photos users take
— which enables the chatbot to perform tasks like image recognition.
Think spotting a specific kind of finch at a bird feeder. https://www.bloomberg.com/news/articles/2024-05-01/ai-startup-anthropic-debuts-claude-chatbot-as-an-iphone-app
Anthropic released an iOS app for their Claude 3 LLM.
I’m past the stage that dismisses LLMs. Some variant will be a useful tool for me. For various tasks. Some I haven’t thought of yet. I’m currently using them as research assistants on topics I’m writing about. To see if detailed prompts (several hundred words with topic headings etc) get responses that include things I’d overlooked. I don’t use any generated text directly.
I might use Claude as a tutor for some studying I plan. #LLM
I think I'll settle on paying for Anthropic Claude 3 via their web interface (I'll check out the API access at some point too), and use PAYG API credits via Drafts for access to GPT 4. The GPT 4 selector in the API currently redirects to gpt-4-turbo.
My first troublesome hallucination with a #LLM in a while: #Claude3#Opus (200k context) insisting that I can configure my existing #Yubikey#GPG keys to work with PKINIT with #Kerberos and helping me for a couple of hours to try to do so — before realising that GPG keys aren't supported for this use case. Whoops.
No real bother other than some wasted time, but a bit painful and disappointing.
Playing around with https://poe.com/ , seriously thinking about quitting ChatGPTplus (paid service) for this, the flexibility in switching models (Claude, Llama, GPT etc) is amazing, I am wondering what I would miss compared with ChatGPTplus. #ai#poe#chatgptplus#llama3#claude
I've recently picked up the habit of asking LLMs a variant of "Is that feature you just told me about real?" immediately after answers that seem too good to be true.
A bit more than half the time my suspicion holds.
"I apologize. I apparently confabulated again."
/cc @driscollis we were just talking about this today.
Does anyone know of a good (preferably free) tool / script that can sort photos into folders for year and month automatically by using the photos’ metadata? Seems like something that should be relatively simple.