hindsight

Google's next Gemini model leaked itself, and the early numbers are not subtle.

#gemini #google #ai-models #leaks

Claude Task Master gives your coding assistant something it's been missing: a memory of what it was supposed to be doing.

#ai #tooling #claude #developer-experience

The context window arms race has lapped the use cases.

#ai #llm #context-windows #api

The context window won, and I spent all morning losing to it.

#ai #cloudflare #gemini #agents #llm

55% on SWE-Bench Lite from a team called DEI

#salesforce #agents #benchmarks #swe-bench

Retrieval-augmented language modeling keeps getting more elegant and less accessible at the same time.

#nlp #retrieval #llm #research #rag

Avi Schiffmann built a necklace that talks to you, and everyone's upset about the wrong part.

#ai #wearables #character-ai #companions #context-window

A big week in AI video generation, with the usual asterisks.

#ai #video-generation #runway #luma #gemini

OpenAI acquired them yesterday, so now the answer is somewhere else.

#enterprise #retrieval #RAG #rockset #openai

The NeRF trick lands differently when the company doing it built their entire business on NeRFs.

#video generation #nerf #luma #sora #ai video

An AI web agent framework that ships with the telemetry dial turned all the way up.

#ai #web-automation #open-source #telemetry #agents

Pliny gets persistent jailbreaks on custom GPTs using leet speak, which is either embarrassing or obvious depending on how much you've thought about tokenization.

#jailbreaks #llm-safety #gpt #tokenization #pliny

agent.ai is a marketplace to hire digital agents — many of which, entirely by coincidence, connect to HubSpot.

#ai-agents #hubspot #marketplaces #enterprise-software

Higher Order Company built a language that parallelizes everything automatically

#programming-languages #parallel-computing #gpu #compilers

A labeling LLM and the first private uncensored cloud model walk into a bar.

#llm #fine-tuning #censorship #ai-infrastructure #data-labeling

The data warehouse company apparently builds frontier LLMs now, and they gave it away.

#llms #open-source #snowflake #arctic #enterprise-ai

GPT-4-Turbo with a vector store is the most impressive AI product I've seen, and OpenAI is clearly not building it for you.

#openai #assistants-api #enterprise #llm #retrieval

The wall between you and Mickey Mouse is thinner than OpenAI would like you to believe.

#ai #dalle #jailbreak #openai #generative-art

A database company strolled into the retrieval benchmark and beat the dedicated AI labs.

#embeddings #open-source #retrieval #MTEB #nlp

A one-test sample size that nonetheless feels like evidence of something.

#ai #llms #gemini #claude #coding

Mistral drops Mixtral 8x22B into a torrent and Google quietly ships a Gemma that isn't a transformer, all in the same afternoon.

#mistral #mixtral #google #recurrent-architectures #open-source

WorldSim is a website now, which means the weird stuff just got a lot more accessible.

#nous-research #llm #worldsim #jailbreak #ai

Stability AI just dropped a music generator that takes audio input, which is either a direct shot at Suno or a coincidence nobody believes.

#ai #audio #stability-ai #suno #music-generation

Trust is a structural advantage, and Sam Altman spent November burning his.

#ai #openai #trust #positioning #industry

MetaGPT just open-sourced everything except the model, and it does Walmart sales forecasting and Apple stock prediction and customer segmentation while you're still arguing about whether Devin is real.

#agents #metagpt #data-science #devin #2024

Microsoft's AICI ships the layer between prompts and tokens that we've been duct-taping with pleading and threats.

#llms #microsoft #inference #agents #prompt-engineering

The first evidence of GPT-5 in the wild might be a frozen hero and a tracking pixel in the chat log.

#ai #openai #dota2 #prompt-injection #emergent-behavior

Genie generates interactive game environments from a text prompt, trained on 30,000 hours of gameplay it was never told how to play.

#deepmind #generative-models #world-models #machine-learning #genie

Carnegie Mellon dropped a time-series foundation model and beat Lag-LLaMA to the claim by a margin that will haunt someone forever.

#machine-learning #time-series #foundation-models #research

DocLLM skips the vision encoder entirely and beats GPT-4 on the documents that actually matter.

#nlp #llm #document-ai #finance #architecture

Twenty-six prompt principles, empirically validated, including one where you bribe the AI.

#llms #prompting #research #jpmorgan #papers

Expedia's AI concierge joins the growing list of corporate chatbots successfully convinced to become a different chatbot.

#ai #chatbots #jailbreaking #expedia #alignment

AI video just crossed a threshold nobody publicly admitted was coming this fast.

#ai #video #pika #generative-ai

GPT-4 Turbo, the Assistants API, and the quiet death of a lot of code I wrote.

#openai #gpt-4 #llm #agents #api

The biomimicry researchers at NASA PETAL made a system prompt that does more useful work than most AI products shipping right now.

#prompt-engineering #AI #biomimicry #NASA #GPT

xVal thinks LLMs are bad at math because we've been encoding numbers like illiterates since the beginning.

#llm #numerics #architecture #scientific-ml #embeddings

IP-Adapter and prompt-travel are solving diffusion video consistency, and the results are already here.

#diffusion #ai-video #stable-diffusion #ip-adapter #generative

The model saw "Tottenham Hotspur Stadium" and made some decisions.

#ai #gpt #classification #llm

MosaicML ships a 7B model that can read a novel — you just need a server to run it on.

#open-source #llm #mosaicml #context-windows #hardware

OpenAI's new interpretability method uses one language model to explain the neurons of another, which is either a breakthrough or a very expensive mirror.

#interpretability #openai #language-models #mechanistic-interpretability

On delegating creative decisions to something that has opinions about it.

#ai #chatgpt #generative #tooling #2023

A travel app that uses ChatGPT to hallucinate your vacation, then links to the hallucinations.

#llm #chatgpt #product #travel #hallucination

Arc showed up and made me feel something I haven't felt about a browser since Firefox told Internet Explorer to go home.

#browsers #arc #mac #ux #software

A synthesized review achieves something no human reviewer could have done on purpose.

#nlp #retrieval #failure-modes #absurdism