📅 Wednesday, July 22nd, 2026 · ⏱️ Read time: 5 min · 🔗 Issue No. 18
You're in the loop — Alibaba strolled into Shanghai's AI conference, dropped a 2.4-trillion-parameter model, and casually declared that only Anthropic's Fable 5 beats it. Meanwhile, the open-weight wave hit a very physical ceiling: Moonshot had to stop taking new Kimi K3 sign-ups because too many people showed up, and its servers ran out of room. The bottleneck on frontier AI, it turns out, is no longer talent — it's electricity.
Today: how to dub a video into another language while keeping your own voice — plus a Wall-Street-style cop for AI, a $188B data company, the frontier that ignores chat, five fresh tools, and a bake-off prompt.
🔁 The Loop
A 2.4T challenger, a $188B milestone, and an AI referee
A new challenger tops the frontier chatter as another mega-model preview lands. Illustration: InTheLoop.
Alibaba previews Qwen3.8-Max — and says only Fable 5 beats it. At Shanghai's World AI Conference on July 19, Alibaba's Qwen team unveiled Qwen3.8-Max-Preview, a 2.4-trillion-parameter sparse mixture-of-experts model with a 1M-token context window, and claimed it ranks "second only to" Anthropic's Fable 5 — with open weights promised to follow. It's live now at 10% of standard pricing on Alibaba's Qoder platform, but there's no model card, license, or third-party benchmark yet. What to watch: whether independent evals and the actual weights back the boast, or whether this joins the growing pile of unverified leaderboard flexes. See the preview.
Databricks is now worth $188 billion — and buying GPUs with it. The data-and-AI platform hit a $188 billion valuation on surging enterprise demand, and says it will spend the fresh capital on scarce GPU capacity to serve customers. Notably, its shares reportedly slipped in after-hours secondary trading — a small sign that even blue-chip AI infrastructure names now face real valuation scrutiny. Read more.
The US floats a Wall Street-style referee for frontier AI. The Trump administration is weighing an independent, industry-funded watchdog — modeled on FINRA and reporting to the SEC — to vet advanced models before or after release, a plan Treasury Secretary Scott Bessent helped shape and now under review by Chief of Staff Susie Wiles. It surfaces the same week Google DeepMind's Demis Hassabis called for a US-led international AI watchdog, sharpening the "does AI need its own FINRA?" debate. The details.
Unlocked reads: The Verge on Google making recipe links prominent in AI Mode, and a fuller rundown of the EU order forcing Android open to rival assistants.
🌊 Deep Current
The AI frontier that isn't a chatbot
The next frontier may be the tables, not the talk. Illustration: InTheLoop.
The billion-euro bet you didn't hear about. While the timeline argued about which chatbot writes the best email, SAP quietly closed its acquisition of Freiburg's Prior Labs and pledged more than €1 billion over four years to scale it — not another language-model shop, but the pioneer of tabular foundation models: AI pretrained on tables and ledgers instead of prose.
Why it matters. Most business data isn't text; it's rows and columns — sales records, inventories, sensor logs. Prior Labs' TabPFN, published in Nature, showed a single pretrained model can beat hand-tuned pipelines on tabular benchmarks with no task-specific training — in one test matching, in 2.8 seconds, an ensemble tuned for four hours.
While everyone debated which chatbot writes the best email, the data that actually runs companies was sitting in spreadsheets no language model reads well.
The catch. The category is young — thinner tooling, fewer integrations, and teams that have spent years bending tabular problems into LLMs out of habit. Adopting a new model shape means retraining people, not just prompts, and SAP is betting a billion euros that the payoff is worth the switch.
The bottom line. If your problem is predicting a number from a table, a tabular foundation model may beat both a fine-tuned LLM and your gradient-boosting pipeline — with far less setup. Watch whether rivals answer SAP's bet before year-end; the quietest frontier in AI may deliver the most measurable ROI.
🛠️ The Workbench
Dub your video into another language — in your own voice
One recording, many markets. Here's how to localize a demo or talking-head clip while keeping your own voice instead of a stock narrator, using a tool from today's Fresh Finds.
Record or upload your source clip — a product demo, webinar segment, or talking-head explainer.
In Cutrix, pick your target languages; it transcribes, translates, and re-times the captions to your pacing.
Turn on "preserve original voice" so the dub keeps your timbre rather than swapping in a generic narrator.
Proof the translated transcript for names, jargon, and numbers — fix these before you render, not after.
Export per language, then A/B the titles and thumbnails; localized copy usually lifts click-through.
Sample Prompt: "Translate this 90-second product demo into Spanish, Portuguese, and German. Keep my voice, match my original pacing, and flag any technical terms or product names you're unsure about before rendering."
🗣️ Overheard
What the timeline's buzzing about
🇨🇳 Kimi capacity crunch: Moonshot paused new Kimi K3 sign-ups after demand maxed out its compute — a reminder that serving a hit model is now a power problem, not just a quality one.
💊 Rubin goes pharma: Bristol Myers became the first drugmaker to buy Nvidia's next-gen Vera Rubin SuperPOD, saying AI already trims molecule-to-clinic time by 20–30%.
🔌 $9.8B for one campus: Hut 8 fully leased its 1-gigawatt Texas data-center site with a second 15-year deal worth $9.8 billion, extending the crypto-miner-to-AI pivot.
🧪 Bezos backs materials AI: CuspAI raised $450M at a $2.6B valuation, with Jeff Bezos and Nvidia in, to build a "search engine" for new materials.
🪙 A $20B transformer chip: Etched is reportedly raising at up to a $20 billion valuation for silicon that runs transformers directly, betting against general-purpose GPUs.
🔎 Fresh Finds
Five tools worth a look
🎬 Cutrix: AI video translation that preserves the original speaker's voice across languages.
🎙️ Universal-3.5 Pro: AssemblyAI's most accurate speech-to-text model yet, via API for voice-driven apps.
🧠 Knowledge Atlas by Fini: a self-learning knowledge base that improves itself over time, aimed at customer-success teams.
🗂️ Compendium: keeps teams, AI agents, and data synchronized on a single shared page.
🖥️ NanoKVM-Go: open-source hardware that gives an AI agent physical control over any screen.
★ = sponsored placement, if any.
🧪 Prompt Lab
The model bake-off scorer
With a new "best" model dropping almost weekly, stop trusting leaderboards and score them on your actual work. Run the same task through two or three models, then paste each output into this prompt.
You are a rigorous evaluation judge.
I will paste the SAME task and 2-3 model outputs labeled A, B, C.
TASK:
"""
[paste the exact task/prompt you gave each model]
"""
OUTPUTS:
A: """[paste]"""
B: """[paste]"""
C: """[paste]"""
Score each output 1-10 on: (1) correctness, (2) completeness, (3) instruction-following, (4) clarity, (5) cost/length efficiency. Show a table of scores. Quote the single strongest and weakest line from each.
Name the winner, state WHY in two sentences, and flag any factual error you caught. If it's a tie on quality, break it in favor of the shorter answer.
Want an image for the write-up? Try this named-style prompt:
A paper-cut diorama of three small robots on a start line facing a finish ribbon, layered pastel paper with soft shadows, restrained indigo-and-amber palette, generous negative space, no text, no letters, no logos.
⏪ Rewind
Yesterday’s most-opened link
The EU's antitrust hammer: the European Commission ordered Google to open Android to rival AI assistants and share search data with competitors, with data-sharing due January 2027 and interoperability by July 2027 — the most consequential AI regulatory move of the year.
Stay in the loop — the InTheLoop team