Nov 28

Nov 21

Temporally stable gen AI video

Nov 6

Released some interesting stuff, GPT4 turbo, 128k context window for GPT4, price * 1/3 on avg. Code interpreter and retrieval, I’m excited to dig into these to see where they netted out. They also released a bunch of features for ChatGPT.

Oct 18

https://lumalabs.ai/

Oct 17

Oct 16

See paper. I take a similar but simpler approach with Edie in that long term memory about the user is uncategorized. As noted in the HN comments this enables graceful degradation of recall that, on it’s surface, seems nice. However it can be off-putting in a chat context because humans don’t tend to forget information in the same way.

The key is to model how humans handle this gracefully degradation, either through traditional means or by training.

Oct 15

Repo author did a great job providing answering why he believes these are not hallucinations.

Oct 14

Supabase puts out a ton of high quality content on LLM app building blocks. Love their preference of battle tested infrastructure. Seems like PGVector is coming along and will be able to handle large scale.

I wonder if the “you’re not web scale” adage applies here. Yes if one or few embeddings per datum, no if 10s or 100s of embeddings per datum.

Oct 2

Video, posted by Alexandr Kurlin to the YC AI/ML group. Succinct one layer down explanation of LLMs.

Sept 20

Landscape overview from Sequoia. Lots of interesting stuff here.

“Some of the best consumer companies have 60-65% DAU/MAU; WhatsApp’s is 85%. By contrast, generative AI apps have a median of 14% (with the notable exception of Character and the “AI companionship” category).” — This makes sense to me based on exploratory work on Edie. Part of the reason is personification of the LLM. Your brain participates in the process similar to rubber ducking.2

Sept 19

Sept 19

Answer: maybe.

Sept 18

Not directly related to generative AI, but the more I learn about FDB the more I think it’s worth a serious look at building the next generation of technology on a more sound footing.

Sept 13

https://simonwillison.net/2023/Aug/3/weird-world-of-llms/

July 20

I mean, it’s not a great episode but that’s not really the point.

July 11

July 7

Deepfake actors into entire shows. As one commenter put it — “The future of movie streaming: please select movie, please select actor.”

June 25

Fantastic resource, see linked PDF. Covers much of the history of

June 18

June 10

None of these people are real.

June 10

June 7 — Warning, mildly NSFW

May 31

May 31

Firsthand account of trying to build an Radiology AI system. First FDA-approved. HN thread here.

May 26

Autonomous agents, different approach than AutoGPT etc. They build up a library of generated code actions and create embeddings of the doc strings. Thoughts → actions → vector search for code actions.

Begs the question, has the 3d simulation that births AGI already been created? Is it Minecraft?

May 21

May 17

May 14

“promptfoo helps you tune LLM prompts systematically across many relevant test cases.”

May 14

May 14

May 13

Raw transcript between Simon and GPT4. Raw transcripts are fun as you get to see how other people interact with the LLM, pick up tips, etc.

As for the content, it reinforces my belief that current LLMs are really good sounding boards, but you need to be an expert in the subject matter to use them effectively (for creating new things, not true for learning, etc).

Meta — May 10

Multimodal model, combine images, text, depth maps, motion data, etc and generate between each modality.

OpenAI project.

Thoughts

Guiding questions / interests:

Declarative approach to AI and non-AI system composition. Where in an imperative system like LangChain you explicitly define program flow, in a declarative system like EVA you define what you want and the system figures out how (I’m assuming).

EVA describes itself as “supporting database applications that operate on both structured (tables, feature vectors) and unstructured data (videos, podcasts, PDFs, etc.) using deep learning models”.

Something to investigate wrt building centralized LLM-based applications.

OP jailbroke Snapchat’s assistant to reveal it’s system message. It’s what you’d expect, but there are some interesting tidbits:

The meta takeaway from this (and ChatGPT jailbreaks) is that the system message shouldn’t be sensitive or secret information — users will find a way to trick the LLM into revealing any information used to generate the response.

From the article: “they understand both the problem (rote cognitive labor) and the solution (large language models).” Fantastic insight that we can use to pattern match whether LLMs are a good solution for a particular problem.

Understanding the real-world use cases where LLMs are better than traditional approaches is important in cutting through the hype, and it tracks that they’re good at rote cognitive labor based on current successes (e.g. Jasper).

Custom-trained LLM for hardware design. Knows the entire context of your project including component list, connections, and related part datasheets.

Eric Michael Smith et al — April 17, 2020

Authors trained a “dialogue manager” to effectively combine existing single-task models (each with existing benchmarks) into a larger blended model. The dialogue manager routed prompts to the best suited single-task model.

While this doesn’t apply to prompting directly, this paper inspired me to consider a similar approach for improving agent responses by implementing a prompt preprocessor / planner to adjust the downstream system message for both ‘act as’ instructions and context facts.

As models grow they typically produce better results for the same prompt except for when they don’t. This prize incentivizes finding prompts that trend worse.

Laird et al — Winter 2017

Overview of different cognitive architectural models.

Lilian Weng — March 2023

Discusses techniques like self-consistency sampling, chain-of-thought prompting, and augmenting (e.g. search engines, APIs) ex: (PAL) and Tool Augmented Language Models (TALM)

Mitchell Hashimoto — April 2023

Founder of Hashicorp explains how to move from ‘try random stuff’ (blind prompting) to a more rigorous method for evolving prompts over time (prompt engineering).

Joon Sung Park et al — April 7, 2023

Use techniques like observation, planning, reflection to create autonomous genai-powered agents. Discusses design of memory that powers defining goals and inner monologue.

April 25, 2023

April 23, 2023

Method for temporal (frame-to-frame) consistency. Extract keyframes from original video, create grid, run through SD, split back to keyframes, ebsynth between, etc. Workflow here: https://www.reddit.com/r/StableDiffusion/comments/11zeb17/tips_for_temporal_stability_while_changing_the/