Aug 12, 2026 · AI & Tech Daily Digest
AI and tech highlights compiled for August 12, 2026, with summaries, links, and commentary.
I. Geopolitics, Regulation & Cross-Border M&A
1. Manus returns to independent operations as Meta deal is unwound; some user data will be deleted (Geopolitics / Regulation)
Summary:
Chinese AI agent company Manus posted “A Note to Our Users” on the evening of August 11, saying it will soon resume operating as an independent company and continue serving millions of users worldwide—closing out Meta’s roughly $2 billion acquisition after Beijing blocked the transaction. Caixin reported on August 12 that prior shareholders including Tencent, ZhenFund, and Sequoia China bought Manus shares back from Meta at about $2 billion. Manus says deleting some data is required to comply with regulations in specific jurisdictions and is not a security incident. Data generated by affected users on or after December 29, 2025 (Meta’s acquisition date) will be deleted between 8:00 a.m. and 7:59 a.m. SGT on August 23–24, 2026; the backup window runs until 7:59 a.m. SGT on August 23, with restoration opening at 8:00 a.m. SGT on August 25. Unaffected users can continue as usual.
Links:
- Manus — A Note to Our Users
- Caixin — Manus cuts ties with Meta as Tencent emerges as top backer
- The Register — Zuck's Chinese agentic prey escapes, will resume standalone ops
Commentary:
The deal became a template for “innovation stays domestic, cross-border data must be erased”—the end state of AI M&A is regulatory restoration, not valuation.
II. Models, Security & Leadership
2. “Stolen Thoughts” research decrypts encrypted reasoning traces; public logs leak keys and passwords (Security)
Summary:
A team of eight researchers from MATS, ELLIS Tübingen, MPI-IS, Snyk, and others published “Stealing Reasoning Traces from Proprietary LLM APIs” (arXiv 2608.09867, submitted August 10) and launched stolen-thoughts.com. They show Anthropic, OpenAI, Google, and others return chain-of-thought as encrypted blocks to clients—and those blocks are portable across sessions, users, and even weaker vs. stronger models within the same provider. Attackers inject a strong model’s encrypted trace into a weak model and jailbreak it to transcribe the frontier model’s full reasoning in plaintext, for roughly two API calls. From 6,708 public agent trajectories on GitHub and Hugging Face they decoded 315,320 blocks and recovered 704 privacy artifacts, including 62 API keys and 33 passwords; they also found instances where Kimi K3 reasoning closely resembles Claude Opus 4.8. GIGAZINE and The Decoder followed on August 12.
Links:
- arXiv — Stealing Reasoning Traces from Proprietary LLM APIs
- Stolen Thoughts — Project site
- The Decoder — Leaked passwords in ChatGPT's hidden reasoning
Commentary:
Hidden reasoning was meant to block distillation—portable client-side ciphertext became a reasoning side channel, and public agent logs are now part of the attack surface.
3. Google DeepMind shake-up: Kavukcuoglu takes day-to-day reins; Hassabis becomes chair and Alphabet chief scientist (Leadership)
Summary:
CNBC reported on August 12 that former DeepMind CTO and Google chief AI architect Koray Kavukcuoglu is now SVP of Google DeepMind, reporting directly to CEO Sundar Pichai and overseeing Gemini model development, frontier AI research, and the Gemini app and developer teams. Co-founder Demis Hassabis stepped down as CEO to become DeepMind chair and Alphabet chief scientist while continuing to lead Isomorphic Labs. Jeff Dean, Sanjay Ghemawat, Oriol Vinyals, and Quoc Le also left to co-found Discovery Loop. The backdrop: no frontier Google release since Gemini 3.1 Pro in February, repeated Gemini 3.5 Pro delays, and key departures including Noam Shazeer and John Jumper to OpenAI and Anthropic; Alphabet shares fell about 5% on the announcement day.
Links:
- CNBC — Google DeepMind: Koray Kavukcuoglu takes over in frontier AI push
- The Next Web — Google named the new boss of Gemini 4. Both of Gemini's co-leads quit the same dayarchived
Commentary:
From “lab CEO” to “SVP reporting to Pichai”—DeepMind is being folded tighter into Google’s product clock; the reshuffle is really a frontier schedule reset.
4. Gemini app passes 1 billion MAUs; Google calls it its fastest-growing consumer product ever (Product)
Summary:
Google CEO Sundar Pichai announced on X on August 11 that the standalone Gemini app (plus web) has surpassed 1 billion monthly active users—the fastest-growing consumer product in Google history and the 14th Google product to hit 1 billion users; the figure excludes Gemini embedded in Search, Workspace, and other channels. Google’s blog said 63% of users have used voice, iOS alone exceeds 100 million MAUs, Gemini generates 150 million+ images daily, and Android users can automate actions across 40+ apps. TechCrunch and The Verge noted on August 11–12 that OpenAI announced ChatGPT also crossed 1 billion MAUs on August 6—both consumer AI front doors now share the “billion-user club.”
Links:
- Google Blog — Gemini app hits 1 billion monthly active users
- TechCrunch — Google's Gemini app surges to 1 billion users
Commentary:
A billion MAUs is a distribution win, not a frontier win—on standalone-app metrics with Android preinstall, Google finally matched ChatGPT on scale.
5. NVIDIA open-sources Nemotron 3.5 Lightning: 30B MoE, 3B active, built for always-on agent execution (Open source)
Summary:
NVIDIA released Nemotron 3.5 Lightning on August 11—a 30B total / 3B active hybrid MoE model (Mamba-2 + MoE + Attention) with up to 1M-token context, aimed at long-running autonomous agents and sub-agent workloads. NVIDIA claims up to ~4× inference speed vs. comparable open models and intelligence on par with gpt-oss-120b-class systems. Weights ship under OpenMDW-1.1 on Hugging Face in BF16 and NVFP4, alongside the NemoClaw open agent security/management stack; SGLang announced Day-0 serving support. Artificial Analysis benchmarks put Terminal-Bench v2.1 at 24%—roughly 3× Nemotron 3 Nano’s 7%.
Links:
- NVIDIA Technical Blog — Nemotron 3.5 Lightning for long-running agents
- Hugging Face — NVIDIA-Nemotron-3.5-Lightning-30B-A3B-NVFP4
Commentary:
The chip vendor is shipping the agent “worker bee” itself—open small MoE plus NVFP4 is the standard play to bind the inference stack to hardware.
6. OpenAI ships Linux preview of ChatGPT/Codex desktop app (Product)
Summary:
OpenAI launched a Linux preview of its ChatGPT desktop app on August 11, unifying ChatGPT, ChatGPT Work, and Codex in one native client. Official support covers Ubuntu 24.04/26.04 LTS, Debian 13, and Fedora 43/44 with x64 and ARM64 .deb and .rpm packages, available globally. The New Stack reported August 11–12 that Linux lacks Native Computer Use at launch (including Appshots, Record & Replay, and desktop-app voice requests) but supports the in-app browser, Chrome extensions, and Codex with local repos and terminals; Anthropic released a Claude Linux desktop app in July.
Links:
- OpenAI Developer Community — Codex in ChatGPT desktop app for Linux is now in preview
- The New Stack — OpenAI's ChatGPT/Codex desktop app is now on Linux
Commentary:
With Mac, Windows, and Linux covered, Codex competition moves from CLI to a persistent workspace—missing Computer Use is a deliberate tier, not an oversight.
III. Funding & Compute Infrastructure
7. Sweden's Lovable raises $400M Series C at $13.3B valuation, backed by EU Scaleup Fund (Funding)
Summary:
Sifted reported on August 12 that Swedish vibe-coding platform Lovable closed a $400 million Series C at a $13.3 billion valuation, co-led by Menlo Ventures and the EU Scaleup Fund—the €5 billion vehicle managed by EQT to back late-stage European strategic companies, with Lovable as a major “application-layer AI” bet. Lovable lets non-developers build apps and sites via natural language; it raised $200M Series A in June 2025 and $330M Series B in December 2025 at a $6.6B valuation. Founder Anton Osika said on LinkedIn the round’s main value is accountability, prioritizing model training, product, infrastructure, and security.
Links:
- Sifted — Lovable raises $400m backed by EU's Scaleup Fund
- TechCrunch — Accel raises $3.5B across four funds to chase AI startupsarchived
Commentary:
EU sovereign capital’s first big bet on “no-code + AI”—the late-stage gap is being filled with breakneck valuations via Scaleup Fund.
8. Skan AI closes $63M Series C, betting enterprise agents need a “context graph of work” (Funding)
Summary:
VentureBeat and Unite.AI reported on August 12 that Menlo Park startup Skan AI raised $63 million in Series C co-led by Cathay Innovation and Dell Technologies Capital, with Citi Ventures, Bloomberg Beta, and others participating, bringing total funding to about $120 million. Skan observes how employees actually work across enterprise software to build a “context graph of work”; it also announced general availability of Skan AI Blueprint and Skan AI Agents alongside existing Skan AI Intelligence—discover, model, automate. Customers include Citi with 75,000 staff.
Links:
- VentureBeat — Skan AI raises $63 million for enterprise context graph
- FinSMEs — Skan AI Raises $63M in Series C Funding
Commentary:
Frontier labs fight for a bigger brain; Skan bets on the work map—agent rollout is missing observability of real workflows, not just models.
9. Indosat, Ooredoo, Nokia, and NVIDIA launch Zankore, targeting 1 GW AI Factory capacity (Infrastructure)
Summary:
A Nokia news release on August 12 said Indonesian carrier Indosat Ooredoo Hutchison, with Ooredoo Group, Nokia, and NVIDIA, launched Zankore by Indosat—a full-stack AI infrastructure platform starting in Asia-Pacific with a 1 GW NVIDIA DSX AI Factory deployment target. About 200 MW is slated for H1 2027 on GB300 NVL72, with DSX MaxLPS dynamic power optimization that NVIDIA says can yield ~40% more effective compute within the same power envelope. Ooredoo is lead investor and platform sponsor; Indosat handles local execution; Nokia supplies AI-native networking; NVIDIA provides accelerated computing and software.
Links:
- Nokia Newsroom — Indosat, Ooredoo, Nokia, and NVIDIA launch Zankore
- EMSNow — Nokia partners with Indosat and NVIDIA to build AI infrastructure in Indonesia
Commentary:
Southeast Asia is copying the “Middle East AI Factory” playbook—telco plus sovereign capital plus NVIDIA reference architecture to lock in regional hub status early.
IV. Enterprise Adoption & Industry Consolidation
10. Ryanair signs five-year Google deal, deploying Gemini Enterprise to 35,000 staff (Enterprise)
Summary:
Unite.AI reported on August 12 that Europe’s largest airline Ryanair signed a five-year Google Cloud agreement to roll out Google Workspace and Gemini Enterprise to about 35,000 employees for crew logistics, operational decision automation, and corporate productivity; Ryanair targets 300 million passengers in fiscal 2034 vs. roughly 216 million today. The deal includes a dual-cloud architecture for critical-service failover and plans to use Google DeepMind’s AlphaEvolve and WeatherNext for fleet ops and maintenance scheduling; crew logistics is named as an early Gemini agent use case.
Links:
Commentary:
Airlines are putting agents first on scheduling and logistics—high-frequency, rule-heavy, low-tolerance work that stress-tests whether enterprise AI can carry operations.
11. Again acquires Genomatica (Geno), merging AI computational biotech with industrial manufacturing (M&A)
Summary:
Tech.eu reported on August 12 that Again acquired industrial biotech firm Genomatica; terms were not disclosed. Geno’s AI and computational biotech platform, patent portfolio, and data spanning experiments, structural analysis, and scale-up will integrate into Again’s computational platform to shorten development cycles for new production pathways and products. Geno focuses on chemicals and materials from renewable feedstocks; Again on feedstock-independent tech and industrial scale-up—combined coverage runs from molecule discovery to factory production, with licensing, co-development, and owned manufacturing as commercial models.
Links:
Commentary:
The next gate for AI-for-biology is whether algorithms connect to fermenters—this M&A locks discovery and scale-up into one P&L.
Today's Summary
- Cross-border enforcement landed: Manus confirmed separation from Meta, data deletion, and shareholder buyback—China’s “restore the pre-deal state” requirement is now operational.
- Security research flashed red: Stolen Thoughts showed encrypted reasoning can be decrypted cross-model; public agent logs are a new leak surface.
- Google’s dual narrative: a DeepMind leadership earthquake alongside Gemini app hitting 1 billion MAUs—reorg chases frontier timelines while consumer scale offsets anxiety.
- Open source and desktop keep expanding: NVIDIA’s Nemotron 3.5 Lightning targets agent workers; OpenAI’s Linux desktop completes the developer triangle.
- Capital and infra: Lovable at $13.3B with EU fund backing, Skan AI on work graphs, and Zankore launching GW-scale AI Factory plans in Indonesia.
Daily Framing:
August 12 was a day when cross-border rules tightened, reasoning-security cracks surfaced, and Google’s leadership reshuffle collided with billion-user scale—geopolitical enforcement and cryptographic architecture flaws running in parallel with the consumer AI arms race and agent infrastructure buildout.
This digest is compiled from real-time search results and is for reference only. Date: August 12, 2026 (Wednesday)