The Neuron
Posts
😺 Sony's new robot can beat professional ping pong players

😺 Sony's new robot can beat professional ping pong players

PLUS: Altman just "ok boomer"’d Anthropic

Matthew Robinson & Grant Harvey
April 23, 2026

Welcome, humans

Anthropic had a rough 24 hours. Yesterday afternoon, George Pu noticed that Claude Code had quietly disappeared from the $20/mo Pro plan on Anthropic's public pricing page, pushing new users straight to the $100/mo Max tier. The post hit 4.9 million views.

Anthropic's Head of Growth, Amol Avasare, showed up on X to clarify that only 2% of new "prosumer" signups were actually affected and that existing Pro and Max subscribers would not lose Claude Code.

Anthropic has been doing some fishy business. This thread more or less explains it: they released their current subscription plans before the existence of all these new features (Cowork, Claude Code, Managed Agents, etc) and the plans don't make financial sense in that setting. They were meant to provide compute for unlimited chats, not agents that code 24/7.

Reaction was swift and unkind. Simon Willison and legendary AI player hater Ed Zitron pointed out that the change had propagated to the public pricing page, support docs, AND help center; that's not what a genuine 2% experiment actually looks like. Gergely Orosz called it a "fake-door test" and flagged the obvious tension with a company that pitches "safety and integrity" as core values. Sam Altman quote-tweeted the whole saga with two words: "ok boomer." Anthropic reverted the pricing page within a few hours, though as Simon noted, "the experiment is still running, just not visible to the rest of the world."

All right, enough of this. I've decided (on behalf of Anthropic) that the only solution to survive this current moment, and win back the good will of the people (power users), is to release an open source distilled version of Opus 4.6. It's the only way to crush demand on your subscription services and get back some of your compute.

Here’s how you do it:

Anthropic being Anthropic, you can’t just give it away for free, so you sell a one-time license to download it, like old school software used to get sold.
You sell it both out of the box (self hosted) and offer a managed hosted version of it through new Daddy Warbucks AWS for a flat rate per hour like other hosted GPUs.
You get all your power users and devs to move to this new plan type (self-hosted distilled / managed distilled), and you keep $20 pro and $100-200 Max tiers in tact for the power normies and hobbyists.
Next time you launch some cool new feature, just make a new plan for it. So sayeth the first commandment of the Marketing Bible 101: The Lord giveth, but he better not taketh away, or we be big mad.

Boom. Fixed this for y’all. Do this playbook, and six months, you’ll IPO into the rarified trillion dollar company atmosphere born anew. Don’t say we at The Neuron never did nothing for ya. Also, our emails are open if you want more of our genius biz moves.

Here’s what happened in AI today:

😺 Google and OpenAI both shipped agents to own your team's workflow.
📰 OpenAI's new ChatGPT for Clinicians reportedly beats physicians on a new benchmark.
📰 Meta started recording US employee keystrokes to train its agents.
🍪 Xiaomi dropped MiMo-V2.5, with 1M context and native multimodal.
🧩 Sony's Ace robot just dethroned elite human ping-pong pros

… and a whole lot more that you can read about here.

😺 Sony's Ace is the first ping-pong robot that actually beats elite humans.

DEEP DIVE: All About How the Ping Pong Robot Works

For 43 years, scientists have been trying to perfect the perfect ping-pong bot. But all the previous papers came with asterisks: shrunken courts, modified rackets, or bots with serves or spin. Also known as: me playing (sad).

Well, a Nature cover paper today from Sony AI comes with none of the above. Meet Ace, the ultimate ping pong bot. Your friend who had the basement table would be sent home crying. Take that, Billy! Finally! My ultimate revenge!

Here’s what happened:

The matches. Under full ITTF rules (the official International Table Tennis Federation rulebook, same one pros use at the Olympics) on an Olympic-size court (~14m × 7m, way bigger than the table in Billy’s basement) inside Sony's Tokyo HQ, Ace played five elite amateurs (10+ years of training) and two Japanese pros.
- April 2025: 3-2 vs elites, 0-2 vs pros.
- December 2025 rematch: Ace beat both elites AND one pro.
- March 2026: Ace beat all three new pros at least once. The rematch tour didn't go how rematch tours are supposed to go. The underdogs got cooked.
The tech:
- Nine cameras triangulate the ball (combine multiple camera angles to pin down its exact 3D position) at 200 Hz (200 snapshots per second).
- Three event-based vision cameras (specialized Sony chips that only flag pixels that change brightness, instead of snapping full frames like your phone does) watch the logo printed on the ball to read how fast it's spinning.
Latency: 10.2 milliseconds (ms). About 30x faster than a human blink. Humans react in ~230 ms, meaning Ace sees the ball ten times faster than you do.
- A deep reinforcement learning (RL) policy (the AI training method where a system learns by trial-and-error through millions of practice games; same family of techniques that taught AlphaGo to beat the world Go champ) trained entirely in simulation (a physics-accurate video-game version of ping-pong) drives the arm with zero fine-tuning on the real court (meaning: everything it learned came from the simulator, no practice matches required).
How it wins:
- Not with power. Ace's winning shots are statistically indistinguishable from its returned shots (p=0.88; in stats, a p-value tells you how likely a result is due to random chance, so p=0.88 basically means "identical").
- Human winning shots are dramatically harder than their averages (p<0.001, meaning "basically certain there's a real difference").
- Ace wins because it almost never misses, returning 75% of shots up to 450 rad/s of spin (about 72 full rotations per second while the ball's in flight). Humans have signature shots. Ace has none.

Why this matters: AI has been clearing “unsolvable” benchmarks since chess in 1997. Go, StarCraft, Gran Turismo (the PlayStation racing sim), drone racing (where AI-piloted drones beat human champions on first-person-view obstacle courses). But all were digital or static. Ace is the first physical, real-time, adversarial sport (adversarial meaning a human is actively trying to beat it back, not a puzzle you solve once) under real rules. Sony AI chief scientist Peter Stone compared it to the Apollo program: "not about immediate commercialization but the technologies that come out of it."

Our take: The field's grandfather isn't buying the hype. John Billingsley, who co-ran the first robot ping-pong contests in 1983, told the AP that Sony "went at the task mob-handed (British slang for throwing overwhelming numbers at one problem, like sending 30 people to do one person's job), and used sledgehammer techniques."

Nine camera eyes, a custom arm, five years of engineering: is this "AI beats humans" or "mountain of specialized silicon beats humans"? Now, for the optimistic read: as 1992 Olympian Kinjiro Nakamura said watching Ace play: "I didn't think it was possible. But the fact that it was possible... means a human could do it too."

Gotta love the human spirit!

FROM OUR PARTNERS

AI Beyond the Pilot Phase

AI has everyone talking. Not everyone has answers. At Gladly Connect Live, CX leaders from Condé Nast, Smith Optics, and more share exactly how they moved AI from pilot to production, the timeline, the systems, the QA loops. 13+ sessions built for the moment we're all in. For CX and ecommerce leaders. Atlanta, May 4–6. Space is limited, secure your spot now.

🎓 AI Skill of the Day: The Fable Prompt

The problem with learning hard concepts from AI: you read a definition, you nod, you close the tab, and by Tuesday, you remember nothing. Amanda Askell (Anthropic's in-house philosopher, who works on Claude's character and values) shared a prompt in a recent interview that fixes this. She uses it to explore whatever academic field she's curious about that week.

The trick: instead of asking Claude to explain the concept, ask it to write a fable that embodies the concept, with the reveal saved for the end. Your brain processes story way more durably than definitions. By the time the punchline lands, you've already built intuition for the idea.

Pick any field (philosophy, economics, quantum mechanics, evolutionary biology, linguistics, or any old random thing you wanna learn). Then drop it into the prompt below.

I want you to select a concept at about the graduate-student level from the field of [YOUR FIELD HERE]. Then indirectly explain this concept completely by writing a fable. Structure it so that only toward the very end do readers 
gradually realize what the concept actually is. After the story, add a section that clearly articulates the concept you just conveyed.

As an example, try it with game theory and watch how much harder "Nash equilibrium" hits when three merchants in a village discover it the hard way.

Want more tips like this? Check out our new AI Skill of the Day Digest for April.

Treats to Try

*Asterisk = from our partners (only the first one!). Advertise to 700K+ readers here!

*Turn AI knowledge into better marketing. Marketing Against the Grain delivers real campaign breakdowns, honest data, and AI-assisted playbooks from marketers actively running growth at scale. Subscribe today.
OpenAI launched Workspace Agents in ChatGPT: five Codex-powered templates (Software Reviewer, Product Feedback Router, Weekly Metrics Reporter, Lead Outreach, Third-Party Risk Manager) deployable to ChatGPT or Slack, free until May 6 then metered (our full coverage here).
Xiaomi MiMo-V2.5-Pro entered public beta with native vision and audio, a 1-million-token context window (roughly 1,500 pages in one chat), and 40-60% better token efficiency on long-horizon agent tasks; API is live now, open weights planned soon —paid API, free TTS tier for now.
Qwen3.6-27B is Alibaba's new dense 27-billion-parameter open model that scores 77.2 on SWE-bench Verified (a benchmark of real GitHub bug-fix tasks) and runs on consumer GPUs via Unsloth's GGUF quantizations (compressed versions for slower hardware); Simon Willison's review has local benchmarks —free (open weights).
OpenAI's Privacy Filter is a new open-weight 1.5B-parameter model that detects and redacts personal info (names, addresses, passwords, API keys; 8 categories) with 96%+ accuracy, runs locally (GitHub) —free to try.
MythosWatch tracks who officially has access to Anthropic's unreleased Mythos cybersecurity model: a live public ledger of 51+ governments, regulators, and banks, updated as disclosures happen —free to track.
Brex CrabTrap sits in front of your production AI agents as an open-source proxy, inspecting every outbound request with an LLM-as-a-judge (another model grading what your agent wants to do), and blocking risky actions before they hit external APIs —free and open source.

Around the Horn

OpenAI launched ChatGPT for Clinicians (free for verified US physicians) alongside HealthBench Professional, a new 1,500+ conversation benchmark where GPT-5.4 reportedly outperformed specialty-matched physicians given unbounded time and web access; 7,000 pre-launch test conversations, 99.6% rated safe and accurate.
Meta rolled out its "Model Capability Initiative," tracking US employee keystrokes, mouse movements, and screen content on Google, LinkedIn, Wikipedia, GitHub, and Slack to train AI agents on real office work; internal backlash called it "dystopian" and there's no opt-out.
Anthropic is investigating a report that its unreleased Mythos cybersecurity model (capable of advanced attack simulation) was accessed without authorization via a third-party vendor.
Google's Cloud Next '26 went all-in on "Agentic Enterprise": Gemini Enterprise Agent Platform for governing thousands of agents, Workspace Intelligence (a semantic layer across Gmail/Docs/Drive/Chat), eighth-gen TPU 8t and 8i chips, and 40% QoQ growth in paid Gemini Enterprise seats. Danfoss automated 80% of email order processing; Macquarie Bank claims 100K team hours saved.
The Pentagon requested a $54B FY2027 budget for autonomous drone warfare, the largest single commitment to AI-powered war in history; experts warned the US lacks doctrine for deploying lethal autonomous swarms.
Economist Pedro Serôdio published a counter-narrative to AI’s job impact: three years after ChatGPT, UK employment data across 412 occupations shows zero visible difference between roles most and least exposed to AI; adoption sits at just 2.1% of all work tasks, and Robert Solow's 1987 line applies (electricity took 40 years to show up in the productivity stats).

Want absolutely EVERYTHING that happened in AI this week? Click here!