Factories of Thought
& The Humanity We Lose Along the Way
I planned to write about the hardcore infra scale-ups – the gigawatt factories of GPUs racing to ten, the sense that we’re building Factories of Thought. You can taste the sci-fi: endless electricity demand, Dyson-Sphere daydreams. Thesis: technology keeps lifting the floor of human possibility; soon capability will be abundant, but judgment remains scarce. If we’re going to build factories of thought, we should decide what those factories are meant to produce.
These thoughts have been percolating my mind as I’ve been bingeing The Expanse lately. If there was ever a show tailor-made to me, this is it: physics with consequences, intergalactic space opera, new worlds, old orders burning. In a highly advanced civilization, a lot of people live on “Basic” (UBI), and without work many wander without purpose. Post-scarcity cures all diseases, solves hunger, but not meaning. That tension sits under everything I’m about to say.
I don’t think technology is a moral agent. The internet wasn’t “good” or “evil.” It was a tool, and we wielded it into both connection and polarization: telemedicine and doomscrolling, open education and outrage engines. Tools widen the feasible set; people choose how to use it.
I say that as someone whose life was widened by a tool. I grew up with no money, often no food, always a month from eviction. But in Salvador, Bahia, what I had were internet cafés. I rode the risky buses at night—tiny girl, 90 lbs soaking wet, a childhood lived with more than one gun pulled to my head—and still went most evenings. I rented a terminal and found a world that hadn’t found me yet. At the far end of that dial-up tunnel was hope: America, need-blind scholarships, free SAT guides. I learned Yale would take a chance on someone like me, and I wrote my college essay on those rented terminals despite the pain and danger that could easily find me there; the undercurrent of a life lived in fear. It was worth it. Fast-forward: I’m a partner at a venture firm, a legal U.S. resident, and my family doesn’t go hungry. Technology didn’t save everyone, but it saved me. Technology elevates and democratizes – I’m living proof.
And yet: are we better as a society post-internet? Healthier, longer-lived—yes. But also more isolated, more inflamed, less nuanced. We mistake disagreement for evil. Our kids are scrolling themselves numb. We’ve lost some shared sense of meaning beyond the self.
So when people ask me whether AI is inherently evil, I give the same answer: no. The right question is whether we are making good designs. Don’t demonize people; interrogate ideas.
Which brings me to the one launch this week that made me viscerally indignant: Meta’s “Vibes.” I’m not going for decel here. I think highly of Meta’s AI work and of the leaders involved. This isn’t a character judgment on them. But it is a design critique. Vibes isn’t “fun AI tools.” It’s an attention apparatus that manufactures stimuli faster than culture, our shared humanity, can assign meaning. Short-form already trains the nervous system to chase high-variance pings; the euphoria of the scroll; the dopamine hit. Vibes removes the very last brake—human effort—and is the antithesis of what we should aspire to build. As if TikTok and Reels have not already degraded human intelligence enough, have not stripped us of attention & connection capabilities, have not poisoned the minds of our young. Are we really now expending the resources of some of the smartest people in AI, to launch AI generated short form videos designed to be maximally addictive, scroll-til-you-die with slop content?
Spinning up factories of thought to make cotton candy – hollow calories, hollow culture – is not my idea of meaning and success. This is why I love sci-fi thought-experiments: it forces first principles. How would you build a new world? If we could start fresh, what do we want our factories to output – disposable dopamine, or durable capability? We are laying gigawatt rails; let’s demand payloads worthy of them: energy breakthroughs, disease models, planetary-scale science, and yes, tools that make people deeper, not just more engaged.
To the Meta team: you’ve carried a lot of open work on your backs. I applaud that. You sit on some of the most valuable human relationship data on earth. You can use it for good, or you can use it for slop.
For the love of God – don’t pick slop.
So I don’t leave ya hanging’ - below for this week’s roundup:
🏭 Datacenters: the mega scale-up
Microsoft’s Fairwater (Wisconsin) is now the canonical example of “AI factory” scale: 315 acres, ~1.2M sq ft across three buildings, “hundreds of thousands” of GB200s, fiber long enough to wrap Earth 4.5×, closed-loop cooling (near-zero operational water), and a second site bringing state spend past $7B. First phase targets early 2026.
OpenAI’s Stargate added five U.S. sites with Oracle/SoftBank, pushing toward ~7 GW planned capacity and the $500B / 10-GW target “by end of 2025.” Oracle says GB200 racks are already arriving in Abilene.
xAI’s Colossus 2 (Memphis) is aiming for the first gigawatt-scale AI cluster; local reporting shows rapid physical progress (119 chillers ~200MW cooling by Aug 22) and even a plan to ship in a power plant. Ambition aside, timelines and power sourcing are the hard part.
👉 The takeway: Power is now product.
🧠 Model releases: speed vs. scope vs. stickiness
Grok 4 Fast (xAI): the play is latency – 2M context, a single stack that flips “reasoning/non-reasoning” at runtime, pitched at real-time voice/search. If xAI wins anywhere soon, it’s fast agentic responses in consumer UX. (note this was technically last week’s release, but I forgot to include - my b, fam!)
DeepSeek v3.1 Terminus: fewer CN/EN mix-ups, better tool use, stronger code/search agents. Good reminder that agent reliability often comes from training pipeline surgery, not bigger weights.
Luma Ray 3: reasoning claims + 16-bit HDR and now in Adobe Firefly. Translation: pipelines, not just prompts. Expect creative pros to adopt faster than enterprises because integration > raw capability.
Meta CWM (Code World Model): a 32B decoder-only, open-weights research release trained on code execution traces and reasoning tasks to study “world models” for code generation. Meta published the report, code, and weights (weights under FAIR’s non-commercial research license). The emphasis is on teaching the model about state transitions and program effects, not just syntax.
It’s a Qwen, Qwen World! Ft: Max / VL / Omni:
Qwen3-Max (and “Thinking” mode) is posting near-frontier scores across SWE-Bench Verified (69.6), Tau2, SuperGPQA, AIME-25—some “heavy mode + tools” claims are near-perfect, which I treat as promising but vendor-adjacent until third-party replications land.
Qwen3-VL-235B-A22B (Apache-2.0) is strong on OSWorld and GUI manipulation (screenshots→HTML/CSS/JS). Open licensing plus capability is why Qwen keeps winning developer mindshare.
Qwen3-Omni (30B MoE, ~3B active): real any-to-any with text/audio/video I/O and streaming; sensible choice for cost-bounded multimodal apps. Qwen
👉 So what:
Differentiation is hard when everyone’s “SOTA.” The sticky wedge right now looks like latency-class + toolability (Grok Fast), open weights + license (Qwen), and distribution into workflows (Ray3→Firefly).
🧑💻 Agents: from vibe to verbs
Two OAI notes worth caring about:
ChatGPT “Pulse”: OpenAI rolled out a Pro-only, mobile feature that starts the conversation (ambient agents, hello!!) each morning with personalized research cards—pulling from chat history and connected apps (opt-in). This is the most concrete step toward mainstream agent behavior in the flagship app.
“Alpha models” blip: brief appearance of “Agent with truncation / prompt expansion” in the model picker hints at systematic work on memory compression and prompt grafting for tool-augmented workflows. Love me a leak. It vanished fast, but direction of travel is clear.
👉 Enterprise take:
Agents don’t become useful because of a new planner – they become useful when context is cheap, tools are reliable, and state is portable across apps. Pulse addresses habit formation; the alpha variants look like the beginning of a memory cost curve shift.
🌐 Platform chess: give agents eyes and facts
Chrome DevTools MCP server (public preview): Pretty sweet launch. It lets AI coding agents control and inspect a real Chrome via the open MCP protocol. In plain English: agents can finally see runtime behavior instead of guessing. Expect a flood of agentic QA, perf, and E2E testing tools.
Google Data Commons MCP server: natural-language access to trusted public datasets (census, climate, etc.) via MCP. If your agent does analysis/briefing, this reduces hallucination pressure and standardizes data fetches.
Microsoft adds Anthropic to Copilot & Copilot Studio: ooooh the relationship between MSFT and OAI ‘aint looking so hot. 💁🏻♀️ New competitor, who ‘dis? Users can now switch models (OpenAI ↔ Anthropic) inside the same workflow. The bigger story is de-risking vendor lock and pushing “model selection as a product feature.” Honestly great for customers. Not great for the further deterioration of OAI <> MSFT vibes.
👉 Net effect:
With MCP + multi-model Copilot, the stack is admitting a future where agents are the UX and models are swappable components behind them. If you sell infra or platforms, your roadmaps should assume pluralistic model catalogs by default.
🤮 *That* Meta Vibes Slop Launch
I already wrote about it up there. I’ve said my peace. So instead, I will share everyone else’s feedback on it. #Takenoprisoners
🧪 Science corner: materials discovery (the “boring that matters”)
To end us on a high note, I give you Johns Hopkins’ ChatGPT Materials Explorer. It’s an excellent example of where all this should lead: domain-grounded agents tied into physics-based datasets (NIST-JARVIS, Materials Project) that predict material properties in seconds and cut experimental cycles. Now this is a great use of AI!
‘Til next week 🫡
Jess