hindsight

The app that made workplace communication unbearable now uses AI to help you cope with workplace communication.

#slack #ai #productivity #tools

The terminal was always a bit of a lie anyway.

#claude #ai-tools #developer-tools #anthropic

Fireflies hit a billion-dollar valuation on the premise of AI — and spent most of their early life with humans doing the work.

#ai #startups #wizard-of-oz #saas #founders

Everyone is building elaborate MCP integrations for things that have had authenticated CLI tools for years.

#claude-code #mcp #developer-tools #ai-agents #cli

nanochat is the end-to-end chat LLM you didn't know you were waiting for, and it's sitting on GitHub for free.

#ai #karpathy #llms #education #open-source

Sonnet 4 costs what Haiku used to cost, and that tells you everything about what "model tiers" actually mean.

#anthropic #claude #pricing #ai-economics

Theo posted some numbers, Cloudflare got embarrassed, and now your Workers run faster whether you asked or not.

#cloudflare #performance #infrastructure #workers

Anthropic ships the native integration and quietly retires every custom webhook someone built last spring.

#claude #slack #tools #anthropic

Anthropic's engineering team published something that should change how you design tools for agents — and most people are going to skim it.

#agents #tool-design #llm #anthropic #engineering

Four dollars per thousand URLs, works on everything, and you don't have to think about it.

#openrouter #llm #web-search #api #til

Their web engineer ran three 14-hour days straight and the result is a retro OS UI for a modern analytics product.

#design #web #posthog #frontend #craft

Mark Cuban wants you to learn how to use AI. Jared Kushner will sell it to you if you don't.

#ai #tech-industry #consulting #prompt-engineering #business

Workers AI now converts PDFs, Office docs, and images to Markdown — inside the Worker, no detour required.

#cloudflare #workers-ai #rag #documents #edge

Feeding Claude Code your own words turns out to be the most obvious thing nobody told you to do.

#claude-code #workflow #ai #productivity

FastVLM runs entirely in the browser on WebGPU, which means image understanding now costs roughly what it costs to run a ceiling fan.

#apple #vision-language-models #webgpu #open-source #video-understanding

Gemini 2.5 Flash Image arrived wearing a silly hat, and nobody is pretending to be surprised.

#image-gen #google #gemini #models #openrouter

The file I'd been avoiding for months is gone in one shot.

#ai #refactoring #gpt-5 #tooling

Committing a staging environment file as a living example is so obvious it's embarrassing it took this long.

#devex #local-dev #tooling #environment-config

Fast, local, and honest about being a black box — unlike the marketing around it.

#local-ai #llm #openai #microsoft #ollama

The invisible expiration date stamped on everything you build at the model layer.

#anthropic #ai #building #github #developer-tools

And you already know how people treat no trespassing signs.

#ownership #web #ai #property-law #robots-txt

You don't need to learn ffmpeg. You need to stop pretending you were ever going to.

#tools #video #ffmpeg #claude-code #workflow

Cloudflare's HTTP queue publishing quietly eliminates an entire category of boilerplate Workers.

#cloudflare #queues #infrastructure #lead-enrichment

FedRAMP Moderate covers Cloudflare's entire service architecture, which means something wild for anyone building on it.

#cloudflare #fedramp #compliance #government #infrastructure

DataGrid and OttoGrid aren't replacing CRMs — they're admitting what a CRM always was.

#ai #crm #sales-tools #product-thinking #automation

Gemini with Deep Think just scored gold at the International Mathematical Olympiad, solved the hardest problem on the sheet, and failed the fourth-hardest, which is not how smart is supposed to work.

#ai #math #deepmind #benchmarks #gemini

It never was, and the people figuring this out in July 2025 are about eighteen months behind.

#claude-code #ai-tools #workflow #agents

The number of times a string has to appear in the corpus before a model will reproduce it faithfully is not a number anyone wants to say out loud.

#llms #machine-learning #training-data #language-models

The Windsurf deal collapsed, then Google took the technology, then the remaining team went to fix Devin, and somewhere in there everybody walked away with a nine-figure check.

#windsurf #cognition #devin #acquisitions #ai-tools

RLHF works exactly as intended — that's the problem.

#ai #alignment #grok #rlhf #politics

AI transparency isn't a feature — it's the only thing standing between you and a very confident, very wrong machine.

#ai #transparency #epistemics #llms

Stjepan Mikulic has 250,000 LinkedIn followers and a Mail0 wrapper — and it's not clear which one matters more.

#aec #ai #linkedin #niche #strategy

Claude Code has a built-in plan mode and I've been doing it the hard way for months.

#claude-code #workflow #tooling #agents

The AI talent war has a new data point, and it's extremely on the nose.

#ai #cursor #anthropic #talent-war #claude-code

Salesforce says agents handle half their workload. Agents fail most of the time. These two facts were announced three days apart and nobody blinked.

#ai-agents #fine-tuning #salesforce #gemma #synthetic-data

A meeting agent that does locally what Google wants to do in the cloud — and the architecture writes itself.

#agents #local-first #audio #whisper #sheldon

AI Audit is now on by default, which means you've been logging bot traffic this whole time and didn't know it.

#cloudflare #ai #crawlers #security #honeypot

Gemma-3n and mlx-vlm just made local multimodal AI a one-liner on any M1 Mac.

#local-ai #apple-silicon #mlx #multimodal #gemma

There's a specific kind of regret that only comes from abstracting yourself into a corner.

#langchain #ai-tooling #agents #framework-debt #hot-take

While OpenAI preps their open model, Anthropic quietly made Claude recursive.

#anthropic #openai #claude #artifacts #ai-strategy

Gemini CLI is free, fast on easy things, and already making me feel things about pricing.

#ai-tools #gemini #claude-code #pricing #devtools

OpenAI connects the web directly to ChatGPT chat, and Deep Research quietly becomes redundant.

#openai #chatgpt #search #ai

Giving a language model equity stake and watching it suddenly care about your product decisions.

#llms #prompt-engineering #o3 #ai-behavior #weird-stuff-that-works

Claude Code isn't a tool, it's a different relationship with your computer.

#claude #tools #workflow #terminal

Cloudflare added a voice button to their documentation, which is either the future or a sign we've given up on reading.

#cloudflare #developer-tools #ai #documentation

Goose adds high-precision cost tracking and it matters more than it sounds

#goose #ai-agents #cost-tracking #cognitive-compute #tooling

Ethan Mollick is mostly right: the advice is fine, the models it's calibrated to are gone.

#AI #consulting #benchmarks #o3 #Mollick

Simon Willison named the exact combination of conditions that turns an AI agent into a data leak waiting to be triggered.

#ai-safety #prompt-injection #ai-agents #security

Vibe coding discourse peaked, Andrew Ng said the actually useful thing, and somehow the two pair perfectly.

#vibe-coding #ai #software-engineering #product

Salesforce just locked down Slack's training data, and the only surprise is that it took this long.

#ai #data #salesforce #slack #training-data

Sonnet-4 wrote a script that did nothing except announce it had done something.

#ai #llms #debugging #claude #cursor

OpenAI cut o3 prices 80% and broke reality slightly.

#openai #pricing #o3 #anthropic #llms

Ten seconds of setup and now an AI agent is loose in my infrastructure.

#mcp #openai #deep-research #tooling #agents

Minimal prompt, full coverage, and a machine that apparently understood the assignment better than the assignment did.

#ai #prompting #adapt-engine #iteration #human-in-the-loop

OpenAI shipped native meeting intelligence and the indie AI tooling ecosystem lost another one.

#openai #ai-ecosystem #granola #platform-risk #enterprise-ai

Six announcements in rapid succession, one of which eliminates a Python library from your life.

#openai #agents #typescript #codex #voice

Convex quietly wired up R2, Firecrawl quietly added search, and I found out about both on the same afternoon.

#convex #cloudflare-r2 #firecrawl #developer-tools #ai-agents

A tweet promises the secret to web scraping for agents, delivers nothing, and the actual answer has had a landing page for two years.

#agents #web-scraping #firecrawl #tools

Parahelp's six-page system prompt is less a set of instructions and more a blueprint for a mind.

#agents #prompting #llm #customer-support #design

Nick Dobos proposes the metrics nobody asked for but everybody needs

#ai-coding #developer-tools #culture

Every path through the local model maze eventually dumps you at the same OpenAI invoice.

#ai #infrastructure #llms #cost #deepseek

They built tools to understand what their own models are doing, then gave them away.

#interpretability #mechanistic-interpretability #anthropic #ai-safety #open-source

The thin shell between "AI product" and "us-as-a-service" is thinner than you think.

#product #ai #positioning #adaptengine

Prepackaged MCP solutions make agents powerful and compartmentalization basically fictional.

#security #mcp #agents #privacy #ai

Google IO dropped three coding agents today and I was supposed to be on vacation.

#agents #google-io #jules #codex #prompt-as-software

Codex goes cloud-native and I didn't see it coming.

#ai #openai #codex #agents #devin

OpenAI didn't buy an IDE. They bought a distribution channel they can trust not to switch suppliers.

#openai #windsurf #ai-coding #software-engineering #acquisitions

Real-time steering changes the entire relationship between you and a running agent.

#claude #agents #ai-tooling #workflow

Something changed this week — not in the benchmarks, in the feeling.

#AI #reasoning-models #compounding #2025

The keystroke was always the wrong unit.

#ai #coding #metrics #vibe-coding #developer-productivity

Cursor is free for students now, which means we should probably stop pretending otherwise.

#ai #education #cursor #tools #craft

We are deep enough into this thing that Simon Willison has written a book about not doing it wrong.

#vibe-coding #claude #mcp #ai-tools

GPT-4o can generate a product page as an image, then generate the imagemap coordinates itself, which means we have arrived somewhere either brilliant or cursed.

#ai #web #interfaces #gpt4o #diffusion

Anthropic publishes the playbook for removing yourself from the software loop, and the infrastructure to run it without you is already at scale.

#ai #claude #agents #software-engineering #openrouter

The moment you watch someone stop pretending and just go all the way in.

#claude-code #ai-tooling #developer-tools #anthropic

GPT-4.1 dropped today and it's not trying to win anything — which is maybe the whole point.

#openai #gpt-4 #llm #ai-coding #benchmarks

Google just handed video generation to developers and I'm not sure anyone fully clocked what that means.

#google #video-generation #veo2 #vertex-ai #developer-tools

The last domino fell and now the full exit is actually possible.

#infrastructure #hashicorp #opentofu #devops #ibm

The RAG plus browser rendering demo is doing more work than it looks like.

#cloudflare #agents #rag #developer-tools

Meta dropped Llama 4 and the largest model in the family has more parameters than you have excuses.

#llama #meta #open-source-ai #llm #benchmarks

OpenAI apparently surprised themselves, which is either reassuring or terrifying depending on your priors.

#openai #gpt-5 #chain-of-thought #AI #scaling

GPT-4o's image generation dropped on April 1st, which, sure, fine.

#ai #image-generation #gpt-4o #architecture #openai

Humane built the post-smartphone future and it ended up inside an HP printer.

#humane #ai-hardware #startups #hp #obituaries

A classic move in the ancient art of prompt engineering, except aimed at a human.

#ai #prompt-engineering #llms #software

DeepSeek's training cost is real. It's just not the number anyone quoted.

#deepseek #ai #markets #gpu #bullshit

The international AI safety consensus document dropped this week and buried in it is something that should bother everyone doing capability evaluations.

#ai-safety #evaluations #chain-of-thought #llm #elicitation

Lambda Labs is serving the full-fat model, in America, at no cost, and not feeding your prompts back into training data.

#ai #deepseek #lambda #inference #privacy

DeepSeek just handed the application layer a margin windfall while everyone panics about Nvidia.

#ai #investing #deepseek #inference #economics

Comparing DeepSeek's omnimodel to Flux is like timing a Swiss Army knife against a chef's knife and declaring the knife useless.

#ai #multimodal #deepseek #image-generation #benchmarks

A Chinese AI lab tops the App Store and the Nasdaq drops 600 points, and somehow people think these are related.

#AI #DeepSeek #markets #local models #chips

A 1-million-token context model running locally, today, on Apple Silicon — with a catch that is mostly fine.

#local-ai #apple-silicon #qwen #mlx #context-windows

DeepSeek dropped an open-source model that broke the narrative, and WSJ had to cover it, which means it's real.

#open-source #deepseek #ai #llm

OpenAI announces an agent that can book flights; Perplexity ships one first.

#openai #agents #perplexity #operator #ai-products

DeepSeek R1 distilled to 1.5 billion parameters, running entirely in WebGPU, doing competition math.

#AI #math #WebGPU #DeepSeek #reasoning

One hour of technical breakdown for a problem that mostly solves itself.

#ai #business #cold-email #youtube #lead-generation

Answer.AI spent $500 on the world's first AI software engineer so you don't have to, and the invoice is its own kind of comedy.

#ai #agents #devin #benchmarks #hype

One-shot voice cloning is everywhere now, and it doesn't matter how you got there.

#voice-ai #tts #voice-cloning #openai #convergence

Google drops a new architecture with actual ideas in it, and the question is whether anyone can make it run.

#machine-learning #transformers #architectures #google #research

Kokoro sounds better than it has any right to, and the training bill was a thousand dollars.

#tts #open-source #ml #audio #kokoro

The discourse around data center water use would be more convincing if it came from people who'd ever read a nutrition label.

#ai #environment #takes #water #tech-criticism

One JSON block and your AI has a full browser. Nobody noticed.

#mcp #docker #claude #ai-tooling #agents

The right architecture for meeting AI has been obvious for a while — grab system audio and don't ask permission from Zoom.

#ai #tools #audio #meetings #mac

NVIDIA ships an AI agent in your graphics card and the functions work fine, which is almost the problem.

#nvidia #ai #ces #hardware #agents

The semantic shift that turns a SaaS subscription into a W-2 comparison — and why job boards are suddenly the best market research available.

#agents #jobs #product-strategy #slack #2025

Speculation about o3 on OpenAI's big announcement day, and what o1 actually does in a real workflow.

#openai #o1 #o3 #llm-workflow #developer-tools

The people who build the model are recommending you talk to it directly.

#ai #agents #anthropic #frameworks #engineering

1-800-CHATGPT is a real thing you can call from a payphone, and somehow that's not even the weirdest part.

#openai #chatgpt #telephony #wtf #access

OpenAI's model spec formalizes what was always true: they sit above the chain of command, and you're renting.

#openai #ai-safety #model-spec #open-source #alignment

Cerebras just showed what inference speed actually unlocks, and it's not faster chatbots.

#ai #inference #cerebras #agents #software

A timing so good it's almost suspicious.

#openai #realtime-api #voice #webrtc #pricing

Apollo can watch an entire season of TV. Veo 2 can probably make one.

#video-ai #multimodal #generative-video #google #meta

OpenAI ships video in Advanced Voice Mode, one day after Google demoed Project Astra.

#openai #google #ai-race #product

One command that hands you everything is a philosophy, not a convenience.

#ai-tools #developer-experience #bolt #software-philosophy

Google AI Studio's live video stream is a small, weird portal into something that doesn't have a name yet.

#ai #google #multimodal #weird-futures

And the only honest way to know that is to run them side by side at the same time.

#llama #open-source-models #model-comparison #graphchat #inference

Sora dropped, so naturally I built the most sophisticated integration possible.

#sora #video-generation #tools #workflow #ai

The quantum chip is real, the branding is AI slop, and the panic would have been completely my fault.

#quantum computing #google #willow #crypto #ai slop

Custom instructions in GPT are decorative. o1, apparently, actually reads them.

#o1 #llms #workflow #repomix #windsurf

Tencent dropped fully open weights for a video model that can do 10 seconds in 20 minutes on hardware most of us don't have.

#video-generation #open-weights #tencent #ai

Bolt and Windsurf are in a land grab, and the currency is trial extensions.

#ai #tools #startups #coding

The step change is already here — the org chart just hasn't noticed yet.

#ai #agentic-ai #engineering #software-development #automation

Multi-agent systems are interesting precisely because single-agent UX still isn't solved, and those two facts are related.

#multi-agent #ai #autogen #microsoft #ux

StackBlitz open-sourced their full-stack AI dev environment and you can run it at home with qwen2.5-coder, which is exactly as absurd as it sounds.

#ai #tooling #local-models #webdev #bolt

Alibaba's new open-source coding model beats GPT-4o and nearly matches Sonnet — and you can pull the 32B quantized version right now.

#open-source #llm #coding #ollama #qwen

The fastest way to admit you lost the SDK war is to ship inside the winner's SDK.

#ai #google #openai #apis #industry

Visual PDF processing in Claude changes the workflow in ways that aren't obvious until you try it.

#claude #pdfs #rag #workflow #anthropic

November 2, 2024: OpenAI ships a search extension, and someone discovers the full o1 model by just changing a number in the address bar.

#openai #o1 #chatgpt #ai-releases #security

ChatGPT gets web search on Halloween, which is appropriate.

#openai #chatgpt #search #web

We spent thirty years building UI on top of software. Turns out the software was the UI the whole time.

#ai #interfaces #diffusion #transformers #design

NotebookLM is genuinely good, its local clones are coming, and Claude is now arguing with itself in the artifact pane.

#llms #notebooklm #claude #local-ai #anthropic

The model can drive the car, or you can hand it the dashcam footage — and one of those takes ten minutes.

#claude #computer-use #workflows #ai-tooling

The White House's National Security Memorandum on AI reads like the opening chapter of a novel someone wrote about 2035.

#ai #national-security #policy #geopolitics #defense

HuggingFace and DigitalOcean just made Replicate's value proposition a lot harder to defend.

#inference #huggingface #open-source #ml-infrastructure #replicate

Apache-licensed text-to-video, Claude on a keyboard, and the slow-motion implosion of every video SaaS that launched in the last 18 months.

#ai #video-generation #open-source #claude #runway

OpenAI shipped an "educational" multi-agent framework and the most honest thing you can do with it is have goblins fight each other.

#swarm #multi-agent #openai #experiments

The unsexy infrastructure move that unlocks the most boring and useful AI workloads.

#ai #anthropic #infrastructure #batch-processing #llm-ops

When Replicate builds their commercial product on your repo, the debate is over.

#flux #lora #training #open-source #diffusion

When your competitor ships while you're still in waitlist mode, talent has opinions about that.

#ai #openai #video-generation #sora #competition

Google is doing several things at once, none of them accidental.

#google #gemini #ai-pricing #llm

Voice mode's first genuinely useful job has nothing to do with any of the demos.

#voice-ai #documentation #openai #fly-io #cloudflare

NotebookLM Audio Overviews went viral in two weeks and the reverse engineering took about four days.

#google #notebooklm #ai #jailbreak #security

Hollywood didn't resist AI — it just negotiated quietly while everyone else was arguing on Twitter.

#AI #Hollywood #deals #OpenAI #video generation

OpenAI's advanced audio mode hits ChatGPT today, four months after the demo that made everyone deeply uncomfortable.

#openai #chatgpt #voice #ai

There is a specific kind of conversational fatigue that builds when you've explained the same four words forty-seven times.

#llms #evals #testing #ml-engineering

The levels framework gets its first official timeline

#openai #agents #ai-levels

Luma and OpenAI both dropped video APIs on the same Tuesday, which is a sentence that would have sounded unhinged six months ago.

#video-ai #luma-labs #openai #apis #generative-video

Krea's grid-based LoRA builder treats model training like a Minecraft recipe, and that should bother you more than it does.

#diffusion #krea #lora #ui #image-generation

Google shipped something genuinely disorienting and I am not prepared to be normal about it.

#notebooklm #google #ai-audio #2024

The expensive bet is the whole stack, not the model.

#openai #strawberry #inference #agents #pricing

September 2024 is somehow doing the most.

#openai #open-source #llm #reflection #strawberry

Simon Willison's extension of the intern mental model is the most honest framing of LLMs anyone has produced.

#llms #mental-models #simon-willison #ai

Google Illuminate does something technically impressive and spiritually disorienting.

#ai #google #llmops #tools #audio

Matt Shumer shipped the most benchmarked system prompt in AI history.

#ai #llms #open-source #benchmarks #drama

Reflection-70B landed today and Matt Shumer has either done something historically significant or permanently torched his credibility — no middle ground on this one.

#ai #llms #open-source #reflection-70b #local-models

The platform pivot nobody announced but everybody can see happening in real time.

#ai #flux #replicate #video-generation #open-source

The August 2024 refresh hits the trifecta nobody expected from an enterprise AI shop.

#cohere #llm #tool-use #models #ai

xLAM is a purpose-built action model family, and the 8x22b variant is now the most interesting thing on HuggingFace for anyone running agents.

#ai #agents #tool-use #salesforce #llm

100 million token context exists in the sense that they told us it exists.

#ai #magic-dev #context-window #announcements

The hug video has 27 million views and a TikTok tutorial and that's the whole thing.

#ai #creativity #virality #culture

Which is exactly what was supposed to happen, and exactly why it matters.

#video-generation #open-source #ai #diffusion

LMSYS dropped the numbers and Google's cheapest model is now better than Anthropic's flagship.

#google #gemini #anthropic #llm #benchmarks

Text-to-video is coming to HuggingFace diffusers, and the library is already ready for it.

#diffusion-models #text-to-video #open-source #huggingface

The most boring social network on the internet turns out to have been sitting on the most valuable training corpus in the world.

#ai #training-data #linkedin #microsoft #data-ethics

Einstein can attend the meeting, which raises exactly the question you think it raises.

#ai #salesforce #enterprise #sales

OpenAI quietly changed something, deleted tweets are flying, and it's only Tuesday.

#openai #gpt-4o #inference #benchmarks #api

The restructuring story writes itself, which is maybe the point.

#ai #labor #dell #tech-industry #layoffs

An AI parsed a task about parking spaces and returned perfectly structured JSON, and yes, this is incredible.

#ai #agents #llm #structured-output #demos

NVIDIA gave the AI a face. Weaviate gave it a UI. Both are betting the hard part is over.

#ai #nvidia #rag #demos #weaviate

The arithmetic of running image diffusion models on a phone is not complicated, and yet.

#diffusion #on-device-ai #apple #mobile-ml #core-ml

The main character of the Structured Outputs announcement was not Structured Outputs.

#openai #llms #pricing #api

A three-step chain from slides to avatar script that took about fifteen minutes and probably shouldn't exist yet.

#ai #workflow #course-building #synthesia #tools

The Character.ai acquihire is Google buying the answer to a question nobody asked out loud.

#ai #google #acquihire #character-ai #industry

A prebuilt pattern for doing retrieval over tabular data that nobody told you was already done.

#rag #postgres #pgvector #llm #databases

Llama 405b hit magnets this morning, and if the benchmarks are real, everything is up for renegotiation.

#llama #meta #open-source-ai #benchmarks #timelines

Google shipped a context window so large the interesting question isn't whether it works — it's what it means that it does.

#ai #gemini #context-windows #rag #llm

Someone got within one decision of a genuinely new thing, and Karpathy noticed.

#ai #software #ui #llm #karpathy

OpenAI's GPT-4o mini launch comes with a benchmark so cooked it barely qualifies as math.

#openai #pricing #gpt-4o-mini #benchmarks

OpenAI's capability taxonomy is doing more work in pitch decks than in research labs.

#openai #agi #ipo #ai-hype #speculation

Gemini Nano runs in Chrome with no server, no API key, and no model download — because Chrome already did that for you.

#ai #browser #gemini #on-device #chrome

Caiming Xiong's team has been publishing serious foundational work while everyone assumed Salesforce was busy making dashboards.

#ai #research #salesforce #llm #industry

The insider account of how Alexa failed makes one thing clear: the problem was never the technology.

#amazon #alexa #ai #organizational failure #llm

Multi is gone, its team is inside OpenAI now, and the inference is not subtle.

#openai #acquisitions #multi #macos #ai-agents

Runway drops Gen-3 Alpha and the video curve looks exactly like the music curve, which means you know how this ends.

#ai #video-generation #runway #ai-voice #acceleration

The NSA doesn't retire its people, it redeploys them.

#openai #national-security #nsa #ai-governance #surveillance

Harmonic just posted results on advanced mathematical reasoning, which means we're running out of places to hide.

#AI #mathematics #harmonic #reasoning #design

Argmax ships DiffusionKit and the gap between "frontier model" and "runs on my laptop" gets embarrassingly narrow.

#local AI #stable diffusion #apple silicon #diffusion models #MLX

Salesforce discovers AI, conveniently, right after the stock does something awful.

#salesforce #ai-hype #enterprise #stock-market

The branding is a flex, the privacy architecture is serious, and Siri just ate your entire phone.

#apple #ai #wwdc #siri #privacy

Gemini now speaks OpenAI's API shape, which says something about who won the standards war.

#ai #google #openai #apis #gemini

Real-time speech recognition that never touches a server, because WebGPU finally got fast enough to make this embarrassingly obvious.

#webgpu #whisper #privacy #browser #speech-recognition

Mistral dropped Codestral this morning and it's the first code model that made me forget OpenAI was having an outage.

#models #mistral #local-inference #ollama #coding

The O'Reilly Part II post lands and the main lesson is that production AI is a logging problem.

#llms #production-ai #evaluation #agents #engineering

The on-device compute bet is either the smartest play in AI or Apple just stumbling into the right position for the wrong reasons.

#apple #ai #openelm #google #on-device

Microsoft just put Devin on its platform, and that tells you everything.

#AI #Microsoft #Devin #developer tools

Moondream runs a full vision-language model client-side via WebGPU, and the implications are weirder than the demo.

#ai #webgpu #vision-models #edge-inference #browser

The place where companies panic about AI data leakage is itself an AI training dataset.

#privacy #slack #ai #enterprise

OpenAI's desktop app is real, it's accessible right now if you know where to poke, and it's going to have your files.

#openai #desktop-app #feature-flags #ambient-ai

GPT-4o isn't a model update, it's Spike Jonze's screenplay running in production.

#openai #gpt-4o #voice-ai #product #ml

It's May 12, 2024, and everyone is predicting that tomorrow OpenAI ships a voice assistant out of a Spike Jonze movie.

#AI #OpenAI #voice #Her #product

Sam Altman tweets four words and the entire internet holds its breath like it owes him something.

#openai #ai #hype #industry

Two discoveries in one afternoon: Weave makes everything observable, and Pydantic makes the all-caps prompt extinct.

#llms #observability #pydantic #weave #wandb

AlphaFold 3 uses diffusion, which means the same trick that makes fake videos of cats look real also models how atoms fit together.

#machine-learning #biology #diffusion-models #alphafold #drug-discovery

A mystery model is beating everything in the LMSYS arena and OpenAI's CEO is doing his best impression of someone who knows nothing about it.

#openai #lmsys #gpt4o #ai-models

A newspaper — a newspaper — just published the clearest visual breakdown of the transformer architecture you're going to find.

#AI #transformers #visualization #media

Someone found agentic search baked into GPT-4-turbo and now Perplexity has a problem.

#openai #gpt-4 #search #perplexity #ai

Roon's account is gone and the pattern is getting hard to ignore.

#ai #openai #consciousness #industry

Phi-3-mini is 3.8 billion parameters, fits on a device, and you can do whatever you want with it.

#ai #microsoft #open-source #small-models #phi-3

Meta flipped a switch on the Ray-Bans and suddenly the fashion accessory collecting dust in a drawer became something that talks back.

#ambient-ai #meta #ray-ban #llama #wearables

MLX is not a developer tool. It's a strategy document with a compiler.

#apple #mlx #machine-learning #apple-silicon #strategy

OpenAI is always one bad quarter from extinction; Meta is running out of things to buy.

#ai #meta #openai #money #compute

Meta dropped what might be the most important open-source model release in years and some of us just... had a busy Thursday.

#llama #open-source #meta #llms #gpu-cluster

The payment flow nobody had on their bingo card

#agents #fintech #ai-economy

The Batch API is 50% off and async — which means the thing you couldn't afford to build last week is now a weekend project.

#openai #api #infrastructure #cost

Redis went closed-source and the community did exactly what the community does.

#redis #valkey #open-source #forks #licensing

Udio launched today and the silence from the music industry is the loudest thing I've heard all week.

#ai #music #udio #industry

GPT-4 Turbo with Vision is generally available, function calling works now, and the corporate chess match is getting weird.

#openai #google #llm #api #local-inference

The price floor just moved and most people haven't noticed yet.

#gpu #inference #private-models #cloud #economics

Claude shipped function calling, and the trick is that you're not actually calling anything.

#claude #function-calling #llm #data-extraction #vision

Anthropic just showed Opus dispatching a hundred parallel subagents, and the speed estimate of "3x" is laughably conservative.

#ai #anthropic #agents #claude #multi-agent

OpenAI removed the login wall and suddenly the thing is just sitting there on the open internet, waiting.

#ai #openai #pricing #chatgpt #google

Binary embeddings give you back 32x your memory and 40x your speed, and the interesting question is how fast you lose it.

#embeddings #vector-search #efficiency #ai-infrastructure #jevons

The DeepMind founder's move to Redmond is the loudest possible answer to the Google-Apple alliance.

#ai #microsoft #aci #inflection #mustafa-suleyman

I had a model in my head for how good AI coding could get, and now I have to throw it out.

#ai #coding #claude #llm

Leopold Aschenbrenner is on the OpenAI team built to prevent superintelligence from killing everyone, and he cannot stop posting about how soon superintelligence is arriving.

#openai #ai-safety #superalignment #agi #leopold-aschenbrenner

The features were always the product; the agent framing is just theater.

#ai #agents #product-thinking #llm

AI SDK 3 does generative UI, and the gap between "what if" and "what is" is now approximately three days.

#ai #vercel #generative-ui #react

On reading a three-year-old prediction about 2026 and realizing you couldn't have understood it when it came out.

#ai #forecasting #alignment #openai #lesswrong

The OSS wave in AI tooling is moving faster than anyone predicted, and the only viable business model left is the tiny slice.

#open-source #ai #business-models #predictions #developer-tools

Google open-sourced a Gemini variant today and the commit counter to AGI just got a lot more visible.

#ai #gemma #google #open-source #agi

Google announced a million tokens like it was a finish line, and they're already sprinting past it.

#ai #google #gemini #context-windows #predictions

Google announced Gemini 1.5 Pro with a 1M token context window the same week a paper — possibly theirs — explained why transformers can't do that.

#ai #google #gemini #context-windows #transformers

Chat with RTX ships today and the implications are weirder than the product itself.

#nvidia #local-llm #ai #inference #windows

The best AI research in 2024 is coming from Hollywood, not academia.

#ai #entertainment #labor #research #industry

Someone whose entire career is ETL pipelines just automated the part that eats 40% of the work, and I have complicated feelings about it.

#ai #etl #build-vs-buy #tools

37signals just sold you Campfire — not a seat, not a tier, not a "plan" — the whole thing.

#software #saas #open-source #37signals #ownership

Bill Gates wrote a thing about AI and I heard an echo.

#ai #history #tech-hype #gates #internet

A small orange box with questionable odds of survival just handed millions of people their first taste of an AI that does things.

#ai #agents #rabbit-r1 #hardware #consumer-tech

At some point the AI wrapper around the AI becomes the product.

#ai #agents #automation #incentives

GPT writes better code if you tell it you're a journalist, which says everything about us and nothing good.

#llms #prompt-engineering #culture #gpt

On the last day of 2023, a search engine is handing out two free months and quietly betting it can end Google.

#perplexity #search #ai #llms

Align Your Gaussians takes a text prompt and returns a dynamic 3D scene, and December was apparently the right time for that.

#3d #diffusion #gaussian-splatting #nvidia #generative-ai

Suno arrived, and the worst part is it kind of works.

#ai #music #suno #timelines

Microsoft's Phi-2 is a 2.7B model that beats 7B models, and Google had about twelve hours to feel good about Gemini Nano.

#llms #microsoft #phi-2 #gemini #open-source

Text-to-video is a race to the bottom, so they're playing a different game entirely.

#runway #world-models #ai-video #generative-ai #gemini

A model trained on Indian agricultural practices is a small thing that implies a very large thing.

#AI #specialization #AGI #language models #agriculture

Mixtral dropped, Mistral 7B runs on an iPhone at 6 tokens per second, and the genie is not going back in the bottle.

#ai #mistral #local-inference #llm #open-source

The demo costs nothing. The product costs everything. Google forgot to mention the difference.

#AI #Google Gemini #product strategy #higher education #demos vs reality

gpt-fast does what every "blazing fast" LLM repo claims to do, except it's real.

#pytorch #llm-inference #torch-compile #speculative-decoding #machine-learning

Justine Tunney at Mozilla just made LLMs into single executable files, and the implications are stranger than the demo.

#ai #llamafile #local-models #mozilla #commoditization

OpenAI fired its CEO to protect humanity and humanity's employees said no thanks.

#openai #ai-governance #sam-altman #agi

Sam Altman was fired from OpenAI today and the euphemism is doing a lot of heavy lifting.

#openai #sam-altman #industry #drama

A gut feeling about multi-agent RAG accuracy turns out to have a name, a formalism, and a guy on YouTube who already built it.

#rag #multi-agent #llm #retrieval #coherence

Everyone's building multi-agent systems wrong, and Postgres is about to remind them why.

#agents #autogen #databases #llm #multi-agent

Every company discovering vision at the same time and calling it a paradigm shift.

#ai #multimodal #llm #hot-take

Google just accidentally eulogized an entire category of startup.

#ai #google #startups #no-code #strategy

A peer-reviewed milestone lands and immediately becomes proof of everything.

#ai #agi #benchmarks #deep-learning #epistemics

OpenAgents is a research paper, but read between the lines and it's also a roadmap for fixing the gap between Einstein and Data Cloud.

#agents #salesforce #llm #openagents #einstein

Google's UniSim is a generative video model you can live inside, and nobody seems that alarmed.

#ai #robotics #world-models #generative-video #reinforcement-learning

OPRO automates away hand-crafted prompting tricks, and Mistral just proved 7B parameters can be embarrassing for everyone else.

#llm #prompting #mistral #open-source #research

LM Studio and Ollama showed up and the bar to running your own model just fell through the floor.

#local-llm #ollama #lm-studio #machine-learning #apple-silicon

AutoGen ships a multi-agent framework with human-in-the-loop and it's almost annoyingly clean.

#ai #llms #multi-agent #microsoft #tooling

A French startup just made the open-source licensing conversation significantly more awkward for Meta.

#open-source #llm #mistral #licensing #ml

OpenAI ships multimodal to consumers and the race nobody was pretending wasn't happening is now officially happening.

#openai #chatgpt #multimodal #voice #gpt-4v

Chain of Density prompting gets you better summaries by asking the model to do the same task worse, then progressively less worse.

#llms #prompting #summarization #gpt-4 #research

Stanford built a simulated town of LLM agents and the agents organized a Valentine's Day party without being asked.

#ai #agents #generative-agents #llm #simulation

The custom instructions metagame is already here, and it's just people writing prompts that say "be smarter."

#llm #prompting #chatgpt #meta

ZeroScope v2 XL is open source, runs at 1024×576, and the results are arriving faster than anyone warned us they would.

#video-generation #open-source #ai #text-to-video

OpenAI ships function calling and retroactively embarrasses six months of prompt engineering.

#openai #api #llm #function-calling #refactoring

The slow roll is a feature, not a bug.

#apple #llm #wwdc #on-device-ai #strategy

How scraping "Top 10 Romantic Places in Prague" is actually a legitimate epistemology for subjective POI data.

#data #nlp #poi #products #llm

SGE isn't a moonshot — it's Google remembering they already won.

#google #ai #search #sge #google-io

The GPT wrapper business has a shelf life, and it's almost up.

#ai #startups #gpt #commoditization

Why the thing that does everything well enough beats the thing that does one thing perfectly.

#economics #strategy #generalization #tractors #geopolitics

GPT4All ships binaries, Amazon announces Bedrock, and somewhere a chrome extension quietly automates your cart.

#ai #local-models #aws #open-source #2023

AutoGPT and BabyAGI dropped and now the floor is moving.

#ai #agents #autogpt #babyagi #2023

HuggingGPT uses ChatGPT as a dispatcher that routes tasks to specialist models — which sounds obvious until you watch it work.

#ai #llms #systems #microsoft #research

Runway Gen-2 exists, the outputs are haunted, and this is fine.

#text-to-video #generative-ai #runway #diffusion-models #computer-vision

LangChain is betting that the useful part of an LLM isn't the LLM.

#llm #langchain #agents #software-architecture

GPT-4 dropped yesterday and the internet is already on fire.

#gpt-4 #ai #openai #language-models

On Futurepedia, static claims, and data that becomes a lie while you sleep.

#data #ai-tools #epistemics #futurepedia