2026-08-14 · 07:08 (CEST)

📱 AI Briefing — 14.08.2026

🚀 Innovation

Qwen3.8 lands: Alibaba's 2.4T MoE tops the open-weight charts. The Qwen team shipped Qwen3.8-2.4T-A95B — a 2.4-trillion-parameter Mixture-of-Experts model with ~95B active parameters — alongside a dense 27B variant and the Qwen3.8-Max flagship, which reportedly claims second place behind Anthropic's Fable 5. Why it matters: open weights are now within striking distance of frontier proprietary quality, though the 2.4T model is a 2TB-RAM-class commitment that most local rigs still can't run.

GLM 5.3 and Muse Glimmer-30B keep the open pipeline full. Zhipu released GLM 5.3, the latest in its MIT-licensed line with 1M-token context, while Meta's Muse Glimmer-30B picked up community quantizations (Unsloth GGUF) within days. OpenAI, meanwhile, previewed an "Ultrafast" mode for GPT-5.6 Sol promising up to 14× inference speed. Why it matters: the open-vs-closed race is now being fought on capability, cost, and speed at the same time — and the open side is not slowing down.

🔬 Research

"Selective activation sparsity" lets smaller models punch above their weight. Presented at ICML 2026, the method trains models to activate only the most relevant parameters per task, matching models roughly three times their size on reasoning benchmarks. Why it matters: it's a credible path to dramatically cheaper training and inference — and to capable models that run on phones and laptops.

Autonomous-agent horizons are doubling every ~123 days. Per a widely-shared 2026 agent-benchmark analysis, METR's task-horizon metric has accelerated from a seven-month doubling pace (2019–2025) to ~123 days, with Opus 4.6 crossing the 14.5-hour mark. Why it matters: it quantifies how fast agents are gaining real autonomy — the exact curve regulators and safety teams are watching, with week-long tasks projected by late 2026.

Evaluation itself is under scrutiny — and training keeps getting cheaper. The worldproof paper diagnoses where world-model predictions break and shows when pixel-level metrics stop being able to rank models at all, while a solo developer trained a 1B-parameter LLM from scratch on 20B tokens for about $200. Why it matters: benchmark trust is the field's bottleneck, even as the cost floor for custom models keeps falling.

🔒 Security

Microsoft's August Patch Tuesday fixes 400+ flaws, including three zero-days. The batch includes critical, unauthenticated remote-code-execution bugs in Windows DNS Server (CVE-2026-62878, CVSS 9.8), DHCP Server, and Active Directory Certificate Services, plus a 9.6 spoofing flaw in SharePoint Online. Why it matters: several are network-facing with no user interaction — this is a patch-now month.

A Windows 11 kernel hole lets attackers skip physical access. Researchers detailed a kernel use-after-free (CVE-2026-62708) affecting Windows 11 24H2/25H2/26H1 and Server 2025 that undermines a standard mitigation. Why it matters: it's a fresh reminder that the world's most-deployed desktop OS still carries locally exploitable kernel weaknesses despite years of hardening.

Windows' SMAP is effectively "pre-disarmed." A netsec write-up experimentally confirmed that the kernel's Supervisor Mode Access Prevention is bypassed on normal syscall/IOCTL dispatch paths (RFLAGS.AC=1) — a ~2,900-location architectural compromise Microsoft documented back in 2020. Separately, ERPNext disclosed that its "Document Follow" feature exposed unauthorized data. Why it matters: fundamental kernel mitigations, not just individual CVEs, remain silently weaker than assumed.

💰 Market

A record $510B in H1 venture funding — and 43% went to two companies. OpenAI and Anthropic together pulled in $217B, and AI captured more than 70% of Q2 global startup capital. Why it matters: strip out the mega-rounds and the market looks like an ordinary 2024/2025 — the headline masks extreme concentration, and public investors are now demanding line-item attribution for AI capex rather than rewarding ambition.

AMD bets up to $5B on Anthropic in a 2GW GPU deal. The strategic equity-plus-Instinct-MI450 partnership shows chipmakers buying their way into the model layer, while Alphabet's $195–205B capex guidance and negative free cash flow sent its stock down ~7%. Why it matters: the frontier is increasingly funded by strategic capital from suppliers and hyperscalers, not venture capital — and the Street has stopped rewarding spend without a return story.

Monetization pressure shows up at both ends. DeepSeek announced new pricing, while OpenAI began serving ads on free and Go plans in India and is rolling out premium seats for ChatGPT Business. Why it matters: with confidential IPO filings in the works, both labs are building the margin story public markets will eventually grade them on.

🏛️ Politics

The EU hands down its first AI Act penalties: €47M across three companies. The AI Office fined an HR firm €18M for an unassessed resume-screening AI, a lender €14M for an opaque credit-scoring model, and a retailer €15M for prohibited real-time emotion recognition. Why it matters: enforcement has replaced aspiration — and the AI Office says Q4 will expand to healthcare, finance, and transport.

US rules get specific about autonomous agents. NIST released AI RMF v1.1, adding the first guidance on multi-agent orchestration and long-running autonomous tasks, while a federal executive order confirms anti-discrimination statutes apply to AI systems with full force. Why it matters: agentic AI now has concrete governance expectations in the US, not just principles.

Public trust in AI leaders is cratering as Chinese-model restrictions harden. A new poll finds young Americans don't trust billionaire AI leaders, while community PSAs flag MiniMax's H3 as restricted in the US, EU, UK, and South Korea. Why it matters: the politics of AI is bifurcating into a trust crisis in the West and de facto use-restrictions on Chinese models.

📎 Sources

← Back to Archive