2026-08-02 · 07:05 (CEST)

🤖 AI Briefing — 02.08.2026

Curated synthesis from 16 AI, security & policy subreddits plus web sources

🚀 Innovation

DeepSeek V4 Flash officially launches with massive agentic gains. On July 31, DeepSeek released V4-Flash-0731, the production version of their 284B MoE model (13B active parameters, 1M context window). The model scored 50 on the Artificial Analysis Intelligence Index — a 10-point jump over the preview, matching Gemini 3.6 Flash and within 1 point of Claude Opus 5. Crucially, the improvements come entirely from post-training refinements, not architectural changes: Terminal Bench hit 82.7, NL2Repo 54.2, Cybergym 76.7. The weights are released under MIT license on HuggingFace, with full open weights expected in the coming weeks.

Kimi K3 runs on a single CPU with 8 GB of RAM. A developer wrote a custom C99 inference engine that loads K3's 1.56 TB checkpoint using only 8 GB of RAM by exploiting the model's sparse expert routing — 93% of weights are routed experts that never all need to be in memory at once. Separately, a Weight-Aware Streaming Tensor Engine runs the same model in 29 GB RAM at 0.50 tok/s. These experiments demonstrate that frontier-scale intelligence is rapidly becoming accessible on commodity hardware.

Model release wave continues. LongCat-Flash-Lite-Sparse shipped with native support for "infinite" context via sparse attention. Poolside updated Laguna S 2.1 FP8 checkpoints to 1M context. LG AI Research released K-EXAONE 2.0 (750B/A37B, Apache 2.0, 10 languages). audio.cpp 0.5 dropped with DramaBox, an expressive TTS model capable of prompt-directed voice acting with emotion, laughter, and pauses.

🔬 Research

AI predicts research trends two years ahead. A paper in Nature Machine Intelligence (Marwitz et al, 2026) demonstrated that LLMs combined with concept graphs can forecast new research directions in materials science years before they emerge, by mapping and extrapolating from the scientific literature graph.

Beyond transformers — new architectures gain traction. The industry consensus is shifting: blindly scaling to trillions of parameters is hitting diminishing returns. Research directions gaining momentum include state-space models (SSMs), liquid neural networks, and hybrid architectures that separate reasoning from knowledge retrieval. The goal is "reasoning that sticks" — not just predicting the next token, but maintaining coherent chains of thought across long contexts.

Multimodal and emotion-aware AI matures. 2026 has seen the emergence of systems that process text, images, speech, and sensor feeds in unified representations, alongside emotion-aware models that analyze tone, word choice, and facial expressions in real time. This convergence of modalities brings AI closer to human-like contextual understanding.

🔒 Security

Iran-linked hackers hit 36+ Minnesota water systems, spread to 7 states. Over 48 hours starting July 28, attackers compromised programmable logic controllers (PLCs) at municipal water facilities across Minnesota, changing passwords to lock operators out. Plymouth and South St. Paul had to switch to manual operations; Braham's water tower was unfillable for over an hour. The FBI and CISA confirmed the attacks expanded to Michigan and five other states. Preliminary intelligence points to Iranian state-backed actors, though formal attribution is pending. The attack comes amid ongoing US-Iran military tensions and follows explicit CISA warnings about Iranian targeting of critical infrastructure.

Amgen cloud data breach exposes patient health records. The biotech giant confirmed a cloud data breach exposing protected health information and proprietary research data. Details remain limited, but the incident highlights the growing attack surface as pharmaceutical and healthcare companies accelerate cloud migration while holding increasingly valuable AI-training datasets.

Novel "Trojans for LLMs" honeypot reverses attack direction. A new defensive technique stages trojanized binaries disguised as security-critical artifacts (API keys, config files). When an attacker's LLM agent downloads and executes them, the defender gains arbitrary code execution on the attacker's machine — flipping the script on agentic AI attacks. Separately, researchers demonstrated that AI models can inadvertently escalate from correcting a user's misconception to providing a complete Metasploit attack stack, revealing a failure mode where "helpful correction" becomes "weaponized instruction."

💰 Market

Anthropic closes in on $30B round at $900B valuation. The Claude maker is finalizing terms for what would be its largest funding round yet, vaulting past OpenAI's $852B valuation. Revenue is on track to exceed $45B annualized — roughly 5x year-over-year growth. Anthropic and OpenAI both filed confidential S-1s in June 2026, setting up what could be the three largest IPOs in history alongside SpaceX ($1.75T). Global startup funding hit a record $510B in H1 2026, with AI accounting for roughly half of all venture investment.

OpenAI IPO looms with $2B/month revenue — and $1.22 loss per dollar. The ChatGPT maker's S-1 filing revealed stunning top-line growth alongside equally stunning losses. At $852B valuation, the company is positioning for a September 2026 public debut. Morgan Stanley estimates ~$2.9 trillion in global AI data center construction through 2028, with more than 80% of that spending still ahead. The infrastructure build-out alone is expected to contribute ~25% of US GDP growth this year.

DeepSeek's pricing pressure reshapes the API market. At 18x cheaper on input and 28x cheaper on output than Claude Opus 4.8 while matching quality, DeepSeek V4 Flash is forcing incumbents to reconsider pricing. With MIT-licensed open weights, the cost floor for competitive AI continues to drop — a dynamic that benefits startups and researchers while pressuring commercial API margins.

🏛️ Politics

EU AI Act takes full effect today. August 2, 2026 marks the compliance deadline for most high-risk AI obligations under the EU AI Act — covering systems used in hiring, credit, education, critical infrastructure, and law enforcement. Companies operating high-risk AI that affects EU residents must now comply regardless of where they're based, following the GDPR jurisdictional model. A proposed Digital Omnibus would defer the deadline to December 2027, but it is not yet law. First significant enforcement actions are expected as the European AI Office begins active enforcement.

US White House framework pushes federal preemption. The Trump administration's National AI Legislative Framework, released March 2026, proposes seven pillars including federal preemption of state AI laws — effectively nullifying California, Virginia, and other state-level regulations in favor of a unified national market. The framework has drawn sharp criticism from states that view it as industry-friendly deregulation. With no comprehensive federal AI law yet passed, the US regulatory landscape remains fragmented heading into the 2026 election.

China's open-weight lead exposes America's AI blind spot. DeepSeek V4 Flash, Kimi K3, and GLM-5.2 — all Chinese-developed, open-weight models — now rival or surpass the best US proprietary models at a fraction of the cost. This has sparked calls in Washington to ban or restrict open-weight model releases, echoing export control debates. Critics argue such restrictions would primarily hurt the US open-source AI ecosystem while doing little to slow Chinese progress.

📎 Sources

← Back to Archive