When the Genius Walks Away: Grok Catches Up, Google Loses Its Legends, Meta Returns to Open Source

A late-night data hall corridor, rows of server racks glowing evenly, an engineer walking away from the lit rows toward an unlit door at the far end where a single small lamp has just been switched on

There is an order in which people come to understanding, and each craft has its own mastery.

—— Han Yu, “On the Teacher” (Tang dynasty, c. 802; translated by the author)

What This Episode Is About

TechWave’s episode on 17 August 2026 (EP150) came after a two-week break — host Harry recorded it fresh off a trip to Vietnam, audibly congested, working through a backlog of tech news. Three topics: SpaceXAI’s Grok 4.6 joining the frontier tier, Jeff Dean and Demis Hassabis both stepping back from Gemini at Google, and Meta’s return to open source. He closed with a long segment about the learning platform he and his co-founder spent eight months building, launching 20 August.

What stayed with me wasn’t any benchmark score. It was the thread running underneath: once a field’s technical path has converged, it stops being worth the smartest people’s time to stay. That single idea explains why Google couldn’t hold its legends, why Meta is willing to open its weights, and what an AI course should actually teach.

Key Points

One: the frontier goes from two players to three, but the real variable is price. Grok 4.6 ties GPT-5.6 Sol on the composite intelligence index (61 to 61), with Fable 5 just ahead at 62. Engineers’ hands-on impression is that Grok sits slightly below Sol and clearly below Fable. The cost picture is a different story: on cost-per-task, Fable runs at least four times Grok, while Grok costs about 70% of GPT-5.6. “Slightly weaker, four times cheaper” doesn’t change a leaderboard — it changes a purchasing decision.

Two: the size gap makes this more interesting. Harry’s estimates: Grok 4.6 is around 1.5 trillion parameters and shares a base model with 4.5, which is why the API price is identical. GPT-5.6 Sol might be three to five trillion; Fable 5 is widely guessed at five to ten. Reaching near-parity at a third of the size is a win for efficiency, not scale. And Grok 4.7 reportedly uses a larger base model, has finished training, and ships within weeks — that’s what he’s actually waiting for.

Three: those 900 million unlocked SpaceX shares were a public experiment. Before the IPO, nearly everyone called the valuation excessive; a loud minority called it a scam — insiders dressing up an AI story to dump on retail. On 6 August the first lock-up expired, releasing 900 million shares, roughly 2.5 times the existing float. The stock rose. Since these holders last had liquidity, the market cap had multiplied several times over; given the chance to realise it, most of them chose to keep holding. The people who know the company best voted with their behaviour.

Four: Harry reads the Google departures as “they didn’t give up on Gemini — Gemini stopped needing them.” The direction is settled: pretraining plus post-training, mixture-of-experts architectures, reasoning and agentic reinforcement learning. Everyone is doing roughly the same thing. What’s left is tuning the training recipe, tuning the data, squeezing out efficiency — work that a top-1% engineer with diligence can do well. He adds a vivid detail: Jeff Dean was famously hands-on, personally tuning pretraining hyperparameters.

Five: leaving two years ago would have been a different event entirely. Back then, test-time compute, reasoning, and how to do reinforcement learning were all unsettled — and whoever chose the direction was the critical person. That’s the judgement worth keeping: the damage a departure does depends on how much uncertainty remains in the field. Dean’s new problem is recursive self-improvement — AI training the next generation of AI. His stated reason for leaving rather than starting a team inside Google: fighting for resources and wading through process is slower than starting fresh.

Six: the cost to Google isn’t short-term. Harry’s counterintuitive call is that losing two people who favoured world models and AI-for-science will push Google DeepMind to concentrate even harder on language-model agents — which on a one-year horizon might make Gemini catch up faster. The price is paid on the long-horizon work. His sharpest line: Gemini is on its third generation, so where is Genie 4? World models may be irrelevant to a digital-only general intelligence, but they are exactly what matters for anything that has to move atoms in the physical world. His own position: not selling, not adding, watching for signs that long-horizon research is being dropped.

Seven: Meta returns to open source, with a lovely technical bonus. Zuckerberg published a long essay — written himself, he stressed — arguing AI power should be distributed rather than concentrated, then opened the weights of their strongest model and released a 30-billion-parameter model that, quantised, runs in 24GB of VRAM. The standout is the companion drafter model shipped alongside it: 1.7GB, it guesses 16 tokens ahead which the large model verifies in a single parallel pass, accepting up to the first mistake. Tokens per second roughly doubles with no quality loss. What makes it unusual is that the drafter is a diffusion model that pulls hidden representations out of the big model’s middle layers — it peeks at the large model’s train of thought before guessing. Harry is candid that he reads Zuckerberg as reading the wind: once Meta has a real shot at the frontier, he puts the odds at about 89% that the openness ends.

Going Further

The news says a company is bleeding talent. Should I sell?

This is the most personal question in the episode. The first reaction to “legendary engineer departs” is fear, the second is to look for an analyst’s take, the third is usually to act on feeling. Harry admits he was rattled — he’d once half-joked that if both of these people left, he’d liquidate.

What he actually did was spend two weeks separating “what happened” from “what it means,” changing his mind once or twice along the way. That sequence is worth stealing. Personnel news is hard to read because it bundles a fact (someone left) with an inference (therefore the company is finished), and the market only trades the second one.

Unbundle it, and the question isn’t how brilliant the person is — it’s how much they were still contributing at the margin. The same Jeff Dean leaving two years ago versus now differs by an order of magnitude, not because he got worse, but because the unsolved problems at that seat ran out. So the real question is: is this company’s biggest open uncertainty about research direction, or about execution efficiency? The first needs geniuses; the second needs institutions and resources.

That gives you something testable. Find a public, dated observation point where the two competing stories predict different things. Harry set one: if the script runs the way he expects, the near term is fine; if he starts seeing long-horizon projects quietly dropped, he changes his view. The value of that sentence is that it can be proven wrong. By contrast, “talent is leaving, therefore bearish” — with no statement of what would make you admit you were wrong — isn’t a judgement. It’s a feeling with a justification attached.

Everyone says it’s a scam. Do I dare touch it?

The hard part about online consensus isn’t that it’s wrong — it’s that it becomes the default. You can hear the frustration in Harry’s voice on this one: calling the SpaceX IPO a scam had become the mainstream Western take, to the point where people doing serious work on the company got attacked for it.

The lock-up expiry was a rare dividing line because it converted talk into behaviour. Calling a company a scam online costs nothing. Sitting on a several-fold gain you can finally realise and choosing not to sell costs something. That’s why insider behaviour carries information — not because insiders are smarter, but because they’re paying to express the view.

There’s a line to hold here, though, or this becomes a different kind of blind faith. Insiders hold for many reasons: tax planning, trading-window restrictions, agreements with the company, plain inertia. So the rigorous version isn’t “they didn’t sell, therefore it’s a good company.” It’s “a 900-million-share unlock failed to break the price, which falsifies the specific prediction that the unlock would trigger a wave of selling.” What got refuted is one concrete forecast, not the entire bear case. Those two sentences carry very different weight, and most online commentary jumps straight to the stronger one.

A habit worth taking: next time you see a fierce bull-bear fight, don’t ask who sounds more reasonable. Ask what date and what number the two sides would predict differently on. Find that point, then come back on that day and check the answer. Arguments with no such point are arguments about identity — enjoy them as entertainment.

Fifteen models in six weeks. Do I have to keep up?

Harry says the pace of model releases is wrecking his nervous system: fifteen significant models in six weeks, more than two a week. Then he does something more persuasive than complaining — he admits he hadn’t touched Grok 4.6 in the four days since launch, because every hour he had for writing code went into shipping his own product, and shipping means using the tools you know cold.

His filter: only pay attention to releases that change your cost structure or change your default tool. Everything else, note that it exists and move on. He applies the same logic to courses — last year people paid to learn an automation tool almost nobody uses now; prompting-tips courses expire the moment the underlying model changes. What’s worth learning is the layer that doesn’t expire: how a web page is assembled, how front end and back end talk, what deployment means, why secrets can’t live in the front end, how version control works.

There’s a direct investing parallel. Most news consumption is about who won this week — a layer that turns over constantly. What compounds is the mechanism by which an industry wins. This episode’s example: the model race is no longer purely about who’s smartest, but about who has the better cost structure at a given level of intelligence. That insight still holds a year from now; the leaderboard doesn’t survive the month.

A quick test for which layer you’re learning: write down what you absorbed last week as a single sentence, then ask whether it will still be true in a year. If not, it was information, not knowledge. Information is fine to consume — it just shouldn’t own most of your hours.

One Thing to Take With You

One idea: how important someone or something is comes down to marginal contribution — not reputation, and not the total they’ve contributed in the past.

Jeff Dean didn’t get worse. What shrank was how much more he could create from that particular seat. When the hard problems in a field have been solved, the strongest person there is doing work others could also do — their scarcity is going unused. This runs the same way at work, at home, and in your own life: how much you’re needed isn’t a measure of how good you are. It’s a measure of how much worse things get without you.

An exercise you can do today: write down the three things currently on your plate, and for each one ask — if I vanished tomorrow, how long would this stall? Anything that stalls for a day before someone picks it up is you patching a hole in a system, and it can be handed off. Anything that would stall for three months is where your actual scarcity lives — so how many of your hours does it get right now?

Most people find that the thing they’re busiest with is precisely the thing someone else could do. That’s not a diligence problem. It’s an allocation problem.

Raw Markdown

This article is an educational discussion of investment method. It is not advice to buy or sell any individual security, offers no target prices, and does not analyze any current holding. Investing carries risk; make your own decisions or consult a qualified professional.