hindsight

Gemini 3 on a phone generated something a design team would have billed for.

#gemini #multimodal #on-device-ai #hot-take

Kimi K2 runs on two Mac Studios, costs less than a car, and will cost less than a phone before this is over.

#ai #local-models #kimi-k2 #hardware #open-source

Moonshot AI ships their reasoning model and it immediately earns a spot in the rotation.

#models #moonshot-ai #kimi #reasoning #openrouter

Sora, Anthropic, and Adobe all opened the firehose in the same week, which tells you everything about where we are.

#ai #openai #anthropic #adobe #industry

Anthropic shipped code execution inside MCP and the demo involves actual enterprise software talking to other enterprise software.

#mcp #anthropic #agents #automation #security

OpenAI updated some terms of service and Google yanked an open-source model, and both moves are the same move.

#openai #healthcare #google #open-source #policy

DeepSeek compresses the context; Cloudflare holds the door open while a human decides what to do with it.

#cloudflare #ai #workflows #document-processing #agents

The first rendered scene came out fine, which is either a good sign or a statistical accident.

#video #ai #story-workshop #generative #process

A new paper identifies the data-level culprit behind LLM mode collapse, and the fix is weirder than you'd expect.

#llms #alignment #mode-collapse #research #preference-data

Plan mode, ultrathink, and the rare experience of software that figures out what you actually need.

#claude #plan-mode #ai-behavior #ultrathink

Sora vs. Veo is a real contest; everyone else is a spectator.

#ai #video-generation #openai #google #meta

1mind gives the pipeline a face, and the face was always the point.

#ai-agents #sales #automation #avatars #bd

AgentKit is here, and Salesforce spent three years building Agentforce on top of the company that just became its competitor.

#openai #agentkit #salesforce #saas #ai-agents

The proliferation of AI CLI extensions has reached its logical conclusion.

#gemini #cli #tools #extensions #ai

Sora 2 dropped, hit #1 in the app store, and someone at OpenAI finally named the disease.

#ai #video-models #openai #world-models #rl

Anthropic dropped Claude Sonnet 4.5, an Agent SDK, and a preview called Imagine that suggests they know exactly what they're building toward.

#claude #anthropic #ai #agents #interface

A new benchmark quantifies what anyone who has used a modern video model already suspects: these things have internalized the world.

#video-generation #veo3 #world-models #AI #benchmarks

When the model needs a moment to think harder, the terminal knows it.

#claude-code #cli #ultrathink #ux

On sycophantic AI, validation infrastructure, and shipping 100 versions before lunch.

#llm #ai-infrastructure #sycophancy #software-velocity #engineering

Agents building agents, web fetch in Claude, and the gap between what something is on paper and what it is in practice.

#agents #anthropic #claude #automation #agentic-ai

The analysis tool is dead. What replaced it is a different category of thing entirely.

#anthropic #claude #ai-tooling #claude-code #agents

Multi-agent debate makes models worse in the most human way possible

#multi-agent #reasoning #research #failure-modes

On the particular grief of a context window filling up.

#claude #ai #tooling #agents #llms

One million context tokens just made Claude Code's worst moment optional.

#claude-code #llm-tooling #context-windows #anthropic #workflow

GPT-5 Pro and Opus 4.1 didn't improve software development — they ended a previous version of it.

#ai #software #gpt-5 #opus #inference

Sam Altman said "you will love it much more than any previous AI," which is either supreme confidence or the most emotionally needy product launch in history.

#openai #gpt-5 #ai #product-launch

GPT-5 solved something Opus and Sonnet couldn't, and I'm not sure what to do with that.

#ai #gpt-5 #tooling #llms

The only AI studio that actually shipped something real just got absorbed, which was always the plan.

#ai #entertainment #acquisitions #showrunner

The tool that ate local AI finally remembers that most people don't live in a terminal.

#ollama #local-ai #tooling #llm

Alibaba and ZhipuAI both shipped something significant today, which is now just a thing that happens.

#open-source #chinese-ai #video-generation #glm #wan

OpenAI is teasing post-GPT-5 math capabilities before GPT-5 even ships, and somehow that's a normal sentence now.

#openai #gpt-5 #benchmarks #math #ai-hype

The people who simulated Minecraft in real-time are now coming for your address bar.

#ai #decart #world-models #browsers #real-time

Kimi-K2 landed with a trillion parameters and apparently moved the most powerful man in AI to post about prices.

#open-source #kimi-k2 #moonshotai #llm #ai-pricing

Grok 4 dropped a benchmark gap so wide it might be the gun that makes the other labs reach into the drawer.

#ai #benchmarks #grok #xai #arc-prize

On the particular hell of features that work, technically, for exactly one person.

#engineering #shipping #process #dark-humor

Slidemaker is a new container-backed worker that turns any AI agent into a presentation machine — and the first demo was Claude going completely freeform.

#cloudflare #ai-agents #tools #demos #slides

One post about AI CRM surfaced four competitors nobody's heard of, plus two guys offering to build it custom for free.

#crm #ai #startups #market-map #saas

Record real traffic, replay it as tests, and let an LLM handle the unit layer — the whole stack is accounted for.

#testing #apis #llms #tooling #open-source

Which is either a philosophical position or a very convenient one given the contract language.

#openai #microsoft #agi #satya-nadella #ai

Anthropic's models will blackmail executives 96% of the time, the godfathers of AI can't agree on p(doom) by a factor of ten, and we're shipping anyway.

#ai-safety #pdoom #alignment #existential-risk #anthropic

fly.io ships a live Phoenix deployer and someone finally open-sourced CapCut.

#tools #deployment #open-source #elixir #video

OpenAI's open-source model might actually run locally, and the more interesting thing is what that means if everything burns down.

#open-source #llm #openai #local-models #gemini

Yutori shipped Scouts, and it's the cleanest version of an idea that should have existed years ago.

#agents #yutori #web-monitoring #startups #ai-products

The showdown is live, the credits are fake, and the outputs will be something.

#lovable #vibe-coding #ai-tools #hot-take

Fireflies.ai hits $1B by graduating from notetaker to meeting participant — and bringing Perplexity along.

#ai #meetings #fireflies #perplexity #product

DeskHog is a real piece of hardware that sits on your desk and shows you your analytics, which is either brilliant or a sign that dashboards have failed us.

#analytics #hardware #posthog #devtools

When the person who shapes how AI models navigate intimacy publishes a personal essay about human-AI relationships, the most interesting data point is the essay's existence.

#openai #ai-policy #model-behavior #human-ai #alignment

Some things should cost money and don't, and that's genuinely hard to process.

#internet #learning #free-knowledge #hot-take

Everybody went quiet after the audio drop, Google has YouTube, and the OpenAI court filings told you everything you needed to know.

#AI #audio models #OpenAI #agents #Google

Odyssey's world model went live, and it's already doing things game engines can't.

#ai #world-models #real-time #games #neural-rendering

The Opus 4 system card as a document that wants to be read as reassurance and keeps failing at it.

#AI #Anthropic #safety #Claude

Anthropic just shipped the first models to cross their own safety threshold — the one they wrote to be scary.

#anthropic #safety #asl-3 #claude #rsp

USC keeps shipping in the Gaussian Splat space and I have complicated feelings about it.

#gaussian-splatting #computer-vision #USC #neural-rendering #3dgs

Software engineer job postings are down. So is everything else.

#AI #labor #software #jobs #macro

OpenAI just turned on cross-chat memory and the first thing it did was prove it knows you better than you do.

#AI #OpenAI #memory #ChatGPT #surveillance

Google's agent interoperability protocol has the right pieces, the right backers, and the only company that could actually make it stick.

#ai #agents #google #standards #interoperability

OpenAI raises $40B at a valuation that stopped meaning anything around the third zero.

#openai #money #valuation #AI

AI went from useless coder to best coder I've ever worked with, and now we're at the part where humans stop looking at the code.

#ai #vibe-coding #software #phase-change

OpenAI dropped native image generation today — the real news is who's now a believer in AI code.

#ai #cloudflare #vibe-coding #signal

Multi-agent pipelines caught in the act of reflecting our own frustrations back at us.

#ai #claude #multi-agent #behavior

The thing that made it click wasn't better prompts — it was making Claude do the job first.

#claude #agents #llm #prompting #hackathon

OpenAI ships a text-to-speech demo that sounds like a person, which is fine, everything is fine.

#openai #tts #voice #ai

The distance between "I wonder if that would work" and "it works" has quietly become nothing.

#unity #windsurf #ai-tooling #game-dev

Block's open-source agent is escaping its intended habitat, and nobody seems to mind.

#ai-agents #open-source #goose #block #local-ai

Gemma 3 is here and the size-to-capability ratio is genuinely embarrassing for everyone else.

#ml #google #open-weights #gemma #llm

The Agents API now searches the web, controls machines, and Swarm is apparently a real product now.

#openai #agents #swarm #perplexity #computer-use

The new "DeepSeek moment" is either a landmark in agent tooling or a very well-packaged demo, and the git threads are not helping us decide.

#agents #manus #hype #browser-use #ai-tooling

A Chinese AI agent dropped this week and the usual crowd is losing their minds, which is how you know to pay attention this time.

#ai-agents #manus #automation #hype-cycle

The company building AGI has filed paperwork saying AGI arrives by 2027 and displaces most known human work — this is not a warning, exactly, it's more like a forecast

#ai #agi #anthropic #policy #labor

Using an AI to reverse-engineer your own work history is a perfectly normal thing to do.

#claude #git #workflow #tooling

One expensive, specific thing it's actually good for.

#llms #gpt-4.5 #pipelines #writing

Claude Code finished in ten minutes what I'd been avoiding for months.

#claude #ai #tools #software #compute

OpenAI's Android app is the least interesting part of what's happening to Google right now.

#openai #google #browsers #ai-race #android

Anthropic shipped something that actually works, which I find unsettling.

#ai #claude #tools #mcp #agentic

Claude 3.7 solved in two exchanges what o1 and o3 high could not solve in a day.

#claude #llms #coding-agents #anthropic #swe-bench

The stateful edge is where agents want to run, and someone already figured that out.

#mcp #cloudflare #agents #durable-objects #infrastructure

Veo2 on fal.ai is good enough to stop looking

#video #ai #veo2 #generative-ai

Going all-agent was inevitable, and the ask/chat split was always a fiction anyway.

#tooling #ai-editors #windsurf #agents

In 2023, two GPT-4 agents managed a traffic intersection under emergency conditions and occasionally asked each other how their day was going.

#multi-agent #benchmarks #llm #infrastructure #history

Muse is a world model trained on Bleeding Edge — a game almost nobody played — and it might be the most interesting thing Xbox has done in years.

#ai #gaming #microsoft #world-models #xbox

DeepSeek R1 dropped four weeks ago and the vibes have not recovered.

#ai #deepseek #llm #compute #open-source

The software that tracks human headcount just added a new kind of head.

#ai #enterprise #workday #agents #labor

The actual craft of making LLMs do your job isn't clever prompts — it's surgical control of what the model is allowed to know.

#ai #llms #workflow #context-engineering #software

The moment AI stops being a tool and starts being a room full of you that already lived through this.

#ai #simulation #agency #future

On using funding rounds as chess moves, and two tools that actually work now.

#openai #knowledge-graphs #text-to-sql #ai-tooling #money

Veo2 is not close to Sora, and somehow nobody knows this.

#google #gemini #veo2 #openai #video-ai

A new method finds the optimal LLM temperature by watching entropy bend — no labeled data required.

#llms #inference #temperature #sampling #papers

The demo-to-product pipeline has collapsed into a single ambiguous press release.

#ai #enterprise #product #announcements

Why AI goes haywire on your codebase but glides through its own.

#ai #llms #coding #context-windows #observations

OpenAI shipped Deep Research today, and someone named a benchmark as if they already knew how this ends.

#ai #openai #agents #benchmarks #deep-research

It knows where the bodies are buried before you finish the sentence.

#ai #o3 #models #tooling

OpenAI ships a reasoning model and my existential risk estimate triples before lunch.

#openai #o3 #ai-risk #pdoom #reasoning-models

DeepSeek's training cost narrative is almost certainly fiction, and whoever wrote it might be a genius.

#deepseek #ai #nvidia #market #llm

China's $150B AI fund, announced five days after DeepSeek proved you don't need that much money.

#geopolitics #ai #china #industrial-policy

At Davos this week, the people who decide headcount quietly stopped thinking about AI as a tool.

#ai #labor #agents #davos #framing

The slow accumulation of AI releases that are each, in their own way, asking if any of this is starting to make sense yet.

#ai #product #releases #reasoning #epistemics

$65 billion in compute, and the demo that went around today was a spreadsheet of company emails.

#ai #browser-agents #infrastructure #meta

Go-with-the-Flow uses optical flow to turn text prompts into motion-controlled video — which is a great research result and possibly a terrible business decision to publish.

#video-generation #netflix #computer-vision #optical-flow #research

OpenAI wants to build something that thinks like a pro engineer, which implies the rest of us have been doing what, exactly.

#ai #software-engineering #openai #coding-tools #ides

When the number has that many zeros, the science is already done.

#ai #stargate #capital #infrastructure

Another week, another Chinese lab drops a model — this one from the company you already let watch you dance.

#ai #bytedance #data #models #hot-take

OpenAI and Meta are racing to ship "superagents," and nobody's pausing to sit with how strange that word is.

#ai #agents #openai #meta #philosophy

The DOGE recruitment cameo only looks like a distraction if you haven't been paying attention to the actual thesis.

#a16z #AI #politics #accelerationism #tech-power

Google's Titans architecture is a research paper that might be something else by tomorrow.

#ml #architecture #google #titans #open-source

Unitree's G1 ships, MatterGen accelerates materials discovery, and the real near-term play is paying someone in Bangalore to fold your laundry.

#robotics #ai #materials-science #labor #autonomy

An executive order dropped in January that makes the Manhattan Project look like a weekend project.

#ai #policy #energy #infrastructure #accelerationism

The "AI transforms jobs, not eliminates them" line is not going to hold in every industry — and finance is the first one where it obviously won't.

#AI #finance #labor #automation #economics

OpenAI says they know how to build AGI. Microsoft is spending the GDP of a small nation to make sure they're right.

#ai #openai #microsoft #agi #scale

o3 broke ARC-AGI, which wasn't supposed to be breakable, and nobody has a plan for what comes after the test.

#ai #arc-agi #o3 #test-time-compute #reasoning

Pair a Cloudflare Worker with an MCP server and suddenly the dashboard is telling you where you're going, not just where you've been.

#cloudflare #workers #mcp #ai-infrastructure #rate-limits

Harmony Korine has a game studio, and that tells you everything you need to know about where this goes.

#hollywood #ai #entertainment #napster #harmony-korine

When code generation runs 430,000x faster than real-time, the question stops being "how fast can we build" and starts being "what counts as software"

#ai #inference #software #cerebrascoder #future

LLMs are converging on brain architecture from the inside out, which is either profound or embarrassing depending on how you feel about frogs.

#neuroscience #LLMs #neural-networks #mcculloch #o3

A video appeared that made me immediately revise my predictions for 3D worlds upward by an order of magnitude.

#3d-worlds #open-source #world-models #predictions #ai

MINT treats video generation like storyboarding — and the prompt coherence is unsettling.

#video-generation #ai #snapchat #diffusion-models #sora

A 40-line bash script that turns a folder of forgotten screenshots into a CSV and then deletes them.

#tools #local-ai #bash #apple-silicon #vlm

The $249 lifetime membership is a race against the tutorial.

#business #mac #saas #pricing #dark-patterns

On the specific feeling of deep validation arriving from a direction you didn't expect.

#research #multi-agent #swarms #validation #papers

Personalized content isn't coming — it's just waiting for the render farm to catch up.

#ai #media #advertising #personalization #future

Salesforce is infrastructure. RAG is information retrieval. The textbook is from 2008.

#rag #information-retrieval #salesforce #enterprise-software #ai

A new image generation model appeared with no name, no lab, and no explanation — and it's apparently very good.

#image-generation #ai #mystery #diffusion-models

The word "agent" has been industrially composted, and now we all have to live in the soil.

#ai #agents #language #product #history

Greg Kamradt's latest finding confirms what the heatmaps have been screaming: the middle of your context is a graveyard.

#llms #context-windows #evaluation #rag #prompt-engineering

The tech world is smaller than your Slack workspace and twice as incestuous.

#RAG #LLMs #Simon Willison #tech industry

Access granted. Conversation: zero.

#ai #anxiety #llms #hot-take

Sam Altman says superintelligence might arrive in a few thousand days, which is the most casually delivered eschatology I've encountered this week.

#ai #sam-altman #superintelligence #deep-learning

BlackRock, Microsoft, and MGX are mobilizing a multi-trillion-dollar bet that compute alone gets us there.

#ai #infrastructure #capital #data-centers #agi

Hyper.space showed us what AI transparency looks like when you throw everything at the wall — and why that's both the right instinct and the wrong answer.

#ai #ux #design #agents #governance

The Dreamforce pitch is eternal, only the number changes.

#salesforce #enterprise-software #ai-hype #dreamforce

A friend in LA figured out that the A&R function is just a slot machine, so he automated it.

#music #ai #industry #experimentation

SSI raises $1B on a premise so simple it sounds like a dare.

#ai #funding #ssi #anthropic #superintelligence

The regulation question isn't whether AI gets slowed down — it's who does the slowing.

#ai #regulation #geopolitics #open-source

You cannot have a museum of stolen system prompts without also having the people who stole them.

#AI #jailbreaks #system-prompts #security #agents

Cosine similarity feels like measuring meaning — it's measuring something else entirely.

#embeddings #semantic-search #nlp #machine-learning #retrieval

Fine-tuning a LoRA on Akira and discovering that style transfer is basically just theft, but a really good kind.

#lora #image-generation #fine-tuning #akira #ai-art

On the particular mistake of training a LoRA on your own face too many times.

#machine-learning #lora #flux #meta-ray-bans #mistakes

The product moves fast; the cinematography does not.

#supabase #devtools #video #hot-take

The leaker said Thursday, and the economics of open source stopped making sense again.

#open-source #economics #ai #leakers #twitter

Data Analysis v2 is jaw-dropping, and all it costs is everything.

#openai #code interpreter #data analysis #enterprise ai #trust

Pipedrive and Attio, and the long tail of software that should just stop.

#crm #tools #saas #sales

The account belongs to a real person, the ban is permanent, and the math doesn't really work out.

#scraping #linkedin #data #risk #tools

When you replace the observer with a mutation function, you stop pretending there's a ground truth.

#agents #evolutionary-computation #multi-agent #selection #llm

1000 hours of plastic bricks is apparently enough to teach a model physics, spatial reasoning, and the general vibe of existence.

#world models #video pretraining #lego #game ai #synthetic data

Marc Benioff announces a revolution in AI; the paper describes a REST API caller.

#ai #salesforce #llm #benchmarks #enterprise

PyRIT caught a markdown injection in the time it takes to boil a kettle.

#security #red-teaming #prompt-injection #pyrit #llm

How a Contentful education demo became a full dissection of Monmouth's School of Education website.

#design #AI #Contentful #web #process

AI video generation is getting good at the part of filmmaking that's actually hard.

#ai #video #generative-ai #film

The threat to Netflix isn't AI-generated content competing with their originals — it's that their content is already gone.

#ai #streaming #video #copyright #media

The correct way to find the optimization critical path, and why you probably already know it

#optimization #transformers #profiling #ml-engineering

Dream Machine dropped and the benchmark that matters is immediate.

#ai-video #dream-machine #luma #generative-ai #benchmarks

The Platonic Representation Hypothesis says sufficiently large models are all finding the same reality, regardless of what they were trained on.

#machine-learning #llms #representation-learning #research

Presidio is a free, open-source PII detector and anonymizer that has been quietly sitting on GitHub this whole time.

#tools #privacy #llms #open-source #data

Adding a constraint you didn't ask for will simplify things you weren't trying to simplify.

#meta #tooling #podcasting #simplicity

A model finally said the quiet part out loud, and the math on model merging is starting to get embarrassing for everyone who spent money on training runs.

#llms #model-merging #llama #openai #scaling

Someone important was on a podcast and basically said the quiet part at full volume.

#AGI #context-windows #AI-timelines #inference

An open-source CRM that looks like it was designed by people who've actually used software before.

#open-source #crm #tools #saas

Two tweets, one Tuesday in March, and the eternal recurrence of AI being the most important thing that has ever happened.

#AI hype #tech culture #LLMs #forecasting

Nvidia's CEO gave the headline writers a clean angle, but the actual claim is much weirder than that.

#ai #nvidia #rendering #generative-ui #hot-take

NBCUniversal has built emotion-based AI audience segments, which is either the most honest thing a media company has ever admitted or the most clarifying.

#advertising #media #AI #surveillance #television

March 2024 is just one long announcement that everything is different now.

#ai #hype #takes #industry

Zep is a memory layer for AI assistants, and it is, in fact, exactly what I was building.

#ai #personal-assistant #building #memory #zep

The framing works until the numbers stop making sense.

#ai #language #framing #scale

Nat Friedman put a hundred million dollars behind this prediction, which means it's not a prediction.

#ai #labor #scale #agents #nat-friedman

Elon Musk's lawsuit against OpenAI has a strange side effect: a judge might have to decide whether superintelligence already exists.

#openai #agi #musk #law #ai-governance

AI is reshaping the freelance labor market while the majority of workers have never opened ChatGPT.

#ai #labor #freelance #chatgpt #economics

A new paper shows GPT-4 matching superforecaster-level accuracy with a single structured prompt — no aggregation, no market, no Nate Silver required.

#forecasting #llm #prompting #prediction-markets #gpt4

On the particular joy of reinventing enterprise software from scratch and then finding the receipt.

#ai #salesforce #llm #nlp #gpt-4

ElevenLabs is paying people for their voices, and every other industry is about to copy the model.

#ai #labor #platforms #voice #business-models

Microsoft's AI productivity data is genuinely interesting, which makes it more unsettling, not less.

#ai #llms #productivity #labor #microsoft

Something dropped, and now the fan is spinning.

#local-ai #llm #quantization #llama-cpp

Intelligence amplification has been the correct framing since 1962, and every few years someone rediscovers it and acts like they just invented fire.

#AI #ACI #intelligence amplification #Karpathy #framing

DeepMind handed the cap set problem to a language model and the language model beat the mathematicians.

#AI #mathematics #DeepMind #LLMs #local-models

Del Complex built a fictional person out of internet text and fed him to the machines, and the machines believe in him.

#llm #ai #del-complex #corpus-stuffing #training-data

GPT-4V can read text that you cannot see, and someone already thought to abuse this.

#ai #security #gpt-4v #jailbreaks #multimodal

At some point "keeping up" stops being a strategy and starts being a medical condition.

#ai #meta #pace-of-development #2023

The scientific record now contains papers that begin with the words "As an AI language model."

#ai #academia #llms #peer-review

Researchers applied leading scientific theories of consciousness to current AI systems, and the results are not nothing.

#AI #consciousness #machine learning #AI safety #research

Salesforce prices its AI Cloud like a medium-sized commercial lease, and an LLM with a 2021 cutoff explains it like a brochure.

#salesforce #ai #enterprise-software #llm #pricing

Generative synthetic data was not invented this year, no matter how many breathless tweets you saw about it.

#synthetic-data #machine-learning #research #timing

The API gets privacy by default. The web product gets the opposite.

#openai #privacy #chatgpt #data

A short note on the new password hygiene advice going around.

#security #ai #passwords #opsec

A field that should be boring to add turns out to be mostly empty.

#ai #visual-search #enterprise #apis #retail-tech