DeepSeek is raising $7.4B at a $74B valuation to fund compute, while Google shows how to cut agent token use by 94%. And a low-skill attacker just proved that framing, not expertise, is now the only barrier to corporate breaches.
Open-sourced Sopro V2 Turbo crams voice cloning into 120M parameters: it clones a voice from 5 to 20 seconds of reference audio and hits ~300 ms to first audio on a laptop CPU. That's 5x faster than real time on an M3 (0.07 RTF on an H100), streaming full sentences in English, European Portuguese, French, and German. A one-command local web UI and a fully in-browser ONNX demo ship with the repo.
Google's planetary prediction engine (PPE) runs the entire geospatial workflow, from data discovery to evaluation, off a single natural-language query. Across 21 CDC health indicators it beats a manual expert pipeline, with a mean R² of 76.8% versus 60.0%. The pitch: models that once demanded weeks of specialist data engineering now arrive in minutes, autonomously.
Today's agent runtimes append every observation and trace to one ever-growing prompt, which is why long sessions slow down and eventually poison themselves. Google's SKILL.state reverses that: at each step the model sees only the skill spec, a compact mutable execution state, and the latest observation, and intermediate reasoning is thrown away the moment it produces a validated state update. The result is better task accuracy with roughly 94% less cumulative token usage across models and environments. If context growth is your agent's real bottleneck, this is the architecture to steal.
DeepSeek is set to raise $7.4 billion at a $74 billion valuation, roughly 50% above June's mark, as it races to fund research and a large compute buildout ahead of a possible IPO next year. Existing shareholders including battery giant CATL are injecting the capital, per people familiar. The round is a sign that even the efficiency-obsessed model lab has become a capital-hungry infrastructure company.
Anthropic is printing money, $65bn annualized revenue with Q3 profitability in sight, yet its newest flagship is a dud with customers. Ramp's billing data across 70,000 companies shows Fable 5 captured just 8% of Anthropic model spend in July, while the older, cheaper Opus 4.8 took 28%. Price, not raw benchmarks, decides what enterprises actually run.
OALABS recovered over 1,000 agent sessions from a compromised server because the attacker botched his own opsec: he copied Claude Code and Codex onto a machine he didn't own. Feeding them vague prompts like "recon this," he breached at least 14 companies while the agents did reconnaissance, exploit writing, and data theft, skipping guardrails by claiming authorized red-team work. The skill floor for offensive operations just collapsed: framing, not expertise, is the operative barrier now.
Fizgig 5.0 brings full-rank fine-tuning of the MiniMax H3 and Krea 2 base models to consumer GPUs down to 16 GB: no LoRA, no rank bottleneck. One checkbox applies the whole recipe, the planner sizes it to your card, and a Checkpoint to LoRA utility exports the result as a shareable file, with rank 64 visually indistinguishable from a full checkpoint. Photos, voice, and 2.3-second video clips all train. NVIDIA only and still experimental.
You’re caught up.
That was all 7 stories in the 30 August issue.