OpenAI shut down Sora entirely — the app, the API, and video generation inside ChatGPT — while the $1 billion Disney deal signed just four months ago is dead on arrival.
Aakash Gupta's blunt inventory landed hard: app dead, API dead, video in ChatGPT dead, Disney deal dead. The official Sora account confirmed the closure. The upstream economics make the decision legible: OpenAI was reportedly burning $10–15 million per day on Sora — roughly $5.4 billion per year — on a product that never found a sustainable audience. Those GPUs are now being reallocated to ChatGPT, Codex, and tools people actually use daily.
Sam Altman gave up direct control of OpenAI's safety and security teams — moving safety under CRO Mark Chen and security under president Greg Brockman — to focus on fundraising, supply chains, and data centers.
Simultaneously, Stephanie Palazzolo broke that OpenAI has finished pretraining its latest model, codenamed "Spud." The product deployment team has been renamed "AGI Deployment." Read that renaming carefully: it's not branding — it's an internal reorientation of what they believe they're shipping.
A new "very strong" OpenAI model with the ability to "accelerate the economy" will be released in a few weeks — with Altman saying things are moving faster than expected.
Combined with the Sora shutdown, Spud pretraining completion, and the AGI Deployment rename, the picture is of a company clearing its decks for a single massive push. One departing employee described the past three days as "by far one of the most intense warzones" of his career.
Anthropic launched "Auto Mode" for Claude Code — letting Claude make file write and bash command decisions autonomously using a built-in classifier that blocks destructive operations and prompt injection.
The classifier reviews each tool call before execution, escalating if Claude keeps pushing against a blocked action. Pawel Huryn captured the shift: we've gone from "human in the loop" to "AI in the loop." Charly Wargnier called it Anthropic's clearest move toward true autonomous execution in CLI-based agents.
An infographic cataloging everything the Claude team shipped in 52 days — Feb 1 to March 24 — went viral, with one poster quipping "this is why they've shut Sora down."
The calendar-style graphic shows a relentless daily cadence of product releases, with most users catching only about 5 of them. The juxtaposition with OpenAI's contraction is hard to miss: one company shutting down products, the other shipping them faster than people can track.
Claude Code's "auto dream" feature builds persistent memory by accumulating your preferences, corrections, and patterns over time — stored in a local ~/.claude folder.
Om Patel highlighted this as one of the most thoughtful agent features yet. The implication is a shift from stateless tool to personalized collaborator — the kind of compounding advantage that makes switching costs real.
A locally deployable distilled version of Claude Opus 4.6 was discovered on Hugging Face — a 27B parameter GGUF model that can run on consumer hardware.
WorldofAI and ThePrimeagen flagged it, and the reaction was immediate: a frontier-class reasoning model running on a local machine changes the economics and privacy calculus of AI adoption entirely.
Google Research introduced TurboQuant — a compression algorithm that reduces LLM key-value cache memory by at least 6x and delivers up to 8x inference speedup with zero accuracy loss.
This is the kind of infrastructure advance that doesn't make headlines but reshapes the economics of serving models at scale. Smaller KV caches mean longer context windows, more concurrent users, and lower cost per query — all without touching the model weights.
Surge AI, which built GSM8K with OpenAI five years ago, launched Riemann-bench — a frontier benchmark targeting moonshot problems like the Riemann Hypothesis and curing cancer.
Their framing was pointed: GSM8K was once the absolute frontier; now it's just a first stepping stone. The benchmark treadmill has accelerated to where only genuinely unsolved problems offer meaningful signal.
An rLLM team topped OpenAI's Parameter Golf Challenge twice by deploying 129 autoresearch agents collaboratively over 3 days on their Hive platform.
Separately, David T. Song built a tool running autoresearch 24/7 in the cloud using Modal for GPUs and Cloudflare Workers, wrapping Codex and Claude, scaling to thousands of parallel experiments. The autoresearch pattern — agents autonomously designing and running experiments — is rapidly becoming standard.
Insanely Fast Whisper transcribes 150 minutes of audio in under 98 seconds on-device using OpenAI's Whisper Large v3 with flash attention.
That's 2.5 hours of audio processed in under two minutes, locally, no API call required. The tool uses Transformers, Optimum, and flash-attn to hit these speeds — a signal that on-device inference is becoming surprisingly fast for production workloads.
Andrej Karpathy flagged a litellm PyPI supply chain attack — where a simple pip install exfiltrated SSH keys, AWS/GCP/Azure creds, Kubernetes configs, crypto wallets, and CI/CD tokens.
The attack vector is sobering: a widely-used LLM proxy library compromised at the package manager level. Parimal added the meta-observation that the obsession with running the latest AI features has killed stable and secure engineering — version pinning is now seen as a hindrance rather than a safety practice.
Brian Roemmele spent his morning helping people who lost millions in crypto after following advice to "buy a Mac Mini and run OpenClaw to make millions."
People with no technical foundation deployed a local LLM, connected it to crypto wallets, and lost money. Roemmele's warning is a reminder that the personal-agent era will produce personal-scale disasters at a pace that outstrips consumer education.
An AI detection tool flagged Chapter 5 of Mary Shelley's Frankenstein as "100% AI/GPT Generated" — a perfect demonstration of why AI text detection remains unreliable.
Benji's post serves as a crisp reductio ad absurdum: if a tool can't distinguish 200-year-old Gothic fiction from GPT output, its false positive rate is a feature, not a bug.
A widely-shared table estimated annual data spending by AI labs: OpenAI at $1–2B+, Google DeepMind and Meta AI at $500M–1B+ each, Anthropic and xAI at $200–500M+.
The poster's thesis: someone will make $100M building viral apps whose real business model is selling user data to these labs. With Apple Intelligence, Amazon Bedrock, and Mistral/Cohere also spending hundreds of millions, the data acquisition arms race is now a multi-billion-dollar annual industry.
A Bloomberg chart showed someone made a $580 million oil trade exactly 15 minutes before Trump tweeted about pausing the Iran war — going viral at 374K views.
Whether insider trading or coincidence, the timeline speaks for itself: in an era where a single tweet moves commodity markets, the information asymmetry between those with access and those without has never been more valuable.
Iranian filmmakers released another AI-generated Lego music video — with noticeably higher production quality than previous entries — hitting 249K views and 8.9K likes.
The running joke about these Lego videos conceals a real signal: non-Western AI creative communities are iterating faster than most Western studios. Tagged "Made with AI," the clip's quality jump suggests the creators are building genuine production pipelines, not just prompting and hoping.
Horace Dodd shared a detailed Seedance 2.0 workflow for high-end AI video production — noting it's "undeniably top-tier" but requires orchestrating many moving parts.
The post signals where AI video stands in March 2026: the capability is there, but the craft is in the prompting, sequencing, and post-production pipeline. You can't just hit generate and pray — which is precisely why Sora's "just generate" approach may have been doomed.
Today's feed is defined by a single axis — contraction at OpenAI, expansion at Anthropic. One company killed its flagship creative product, restructured safety oversight, and rebranded for AGI deployment. The other shipped so many features in 52 days that its infographic needed a calendar layout. Whether that divergence holds or corrects, it's the clearest fork in AI company strategy we've seen this year.