Overview
Today’s thread ties three arcs together. First, the AGI curve looks steep, though the bill for compute could be steeper, as Dario Amodei sets out in a widely watched interview. Second, everyday developer life is tilting toward agent-first workflows and real-time multimodal models. Third, video generation jumps from meme to moviemaking with Seedance 2.0, stirring both fandom and legal headaches. Around that, hard tech keeps humming, from a new ISS crew to Starlink’s scale, with a side of kitchen physics and fresh takes on talent pipelines.
The big picture
Amodei on scaling, diffusion, and the road to profit
Dwarkesh Patel’s interview with Anthropic CEO Dario Amodei sketches an aggressive timeline, from “country of geniuses” systems in 1 to 3 years to fast economic spread via pre-training and in-context learning. He is cool on diffusion models for full capabilities, projects Anthropic’s revenue at $9 to $10B before 2030 despite enterprise friction, and argues for US transparency standards plus chip export controls to manage bio risk and geopolitics.
The compute cliff behind AGI optimism
A clipped segment from the same conversation lays out the financial tightrope. If revenue grows 10x yearly to $1T by 2027, firms could still sink under multi-trillion compute commitments if growth undershoots. Amodei argues for hundreds-of-billions scale rather than trillions to avoid unhedgeable risk.
MiniCPM-o 4.5 goes full duplex
OpenBMB’s 9B-parameter model can see, listen, and speak at once, not in turns. Built on Qwen3-8B with SigLip2 vision, Whisper ASR, and CosyVoice2 TTS, it handles live interruptions, narrates drawings, and scans prices while talking. It runs locally via llama.cpp-omni and posts a 77.6 OpenCompass average, with claims of topping flagship closed models on that suite.
How to code with agents without losing your mind
Hesamation argues that managing AI coding agents feels like leading junior engineers. Scope tasks, accept imperfect but moving outputs, and design codebases that agents can navigate rather than imposing personal quirks. The skill is delegation and iteration, not micromanagement.
Cline CLI 2.0 turns your terminal into an agent workshop
Cline’s update brings parallel agents, headless CI/CD, and Agent Client Protocol support so you can keep your editor of choice. Model-agnostic and usable with local LLMs, it is installable via npm and briefly offers free access to models like Minimax M2.5.
x-cli puts X in your prompt loop
Elliot Arledge ships a lightweight CLI for X’s API v2. Post, search, and manage bookmarks from the terminal, then wire it into Claude or other agents without burning context. It installs via uv and authenticates with your developer keys.
Seedance 2.0 lights up anime fandom
A fast-spreading Goku fight clip shows ByteDance’s tool handling motion, energy effects, and character consistency better than past gens. Fans are trading prompts and edits, with a sense that action anime is about to be remixed at home.
From fan fiction to fan cinema
VraserX calls it plainly, showing Attack on Titan scenes turned into live action style. The buzz met instant controversy, with ByteDance pausing real-human cloning features after legal pushback. The excitement, and the nerves, both feel real.
“Even the haters” shared the dog-at-desk short
Seedance 2.0 also lands quiet, emotive beats. A Bichon Frise studying under a lamp closes with a perfect glance, while earlier posts from the same creator leaned into playful kaiju pets. Replies swing between delight and debates over purpose.
Hollywood, meet the prompt era
Yuchen Jin’s Star Wars skit shows 1080p text-to-video with synced dialogue and actor likenesses, underpinned by a dual-branch diffusion transformer that keeps frames consistent. The claim is blunt, that anyone can prompt synthetic actors now, and that this comes from a Chinese lab.
NASA’s busy week in orbit and on the Moon track
SpaceX Crew-12 launched commander Zena Cardman, pilot Jack Hathaway, and mission specialists Sophie Adenot and Michael López-Alegría to the ISS for research on health and space-grown food. Artemis II prep continues with hydrogen tests and repairs, while a sounding rocket mapped auroral currents in Alaska.
Starlink crosses 10 million customers
Starlink says it doubled users in under a year, now serving more than 10M across 160 countries. Testimonials highlight rural links and disaster response, with replies pointing to price pressure and orbital congestion as the network keeps scaling.
Musk’s unseen legacy, according to Katherine Boyle
Dustin highlights Boyle’s claim that Musk’s biggest US contribution is retraining two generations of engineers to build in the real world. SpaceX and Tesla alumni now seed hard-tech firms, with hundreds of Falcon 9 launches as proof that physical engineering compounds too.
Physics in your frying pan
Impulse Labs’ induction cooktop holds the Leidenfrost point so eggs “float” on steam and slide off stainless steel without oil. The clip shows a clean fry and flip, with replies trading notes on browning and flavour trade-offs.
ElevenCreative compresses campaign cycles
At the London summit, ElevenLabs demoed a tool that pulls speech, music, SFX, images, and video into a single workflow, with “Flows” for rapid iteration. Marketers like the speed, while others worry about brand voice and templated outputs.
A free, intense AI fellowship
Austen Allred pitches a 10-week, costs-covered programme in Austin that targets $200k-plus roles. The video shows 100-hour weeks and heavy token spend, drawing both ambition and scepticism in replies.
Why it matters
AGI timelines and solvency can clash. Even bullish leaders are tempering data centre bets because multi-year chip deals can sink firms if demand misses by a modest margin. Expect disciplined capex, creative prepay structures, and policy to loom as large as model weights.
Developer practice is changing fast. Full-duplex models, terminal-native agents, and small CLIs that talk to social APIs point to an agent-first loop. The scarce skill is scoping, review, and system design, not keypress speed.
Video generation has crossed a line. Seedance 2.0 puts studio-grade looks in the hands of fans, which is thrilling for creators and alarming for rights holders. Watermarking, licensing markets, and platform rules will need to evolve in weeks, not years.
Hard tech is back in the spotlight. A fresh ISS crew, Starlink’s global footprint, and Boyle’s take on Musk signal a swing toward factories, test stands, and reliability curves. Software still matters, but metal and tolerance stacks are calling the shots more often.
Consumer physics can still surprise. Holding the Leidenfrost zone shows how tight control unlocks new kitchen tricks, while reminding us that optimisation can trade taste for ease. Expect similar patterns in home devices where sensors meet classical laws.
The talent race is heating up. Free, intensive training promises quick lift, yet outcomes will hinge on placement credibility, sustained learning, and well-being. As AI tools improve, judgement and domain depth will separate standouts from the pack.
