The day's tech, sifted: Aug 13, 2026

What matters today: Anthropic's investors expect the AI lab to float at $2 trillion or more when it goes public in October, which would make it the largest IPO ever, the same day Databricks closed $5B at a $190B valuation and DeepSeek raised its own API prices ahead of a possible listing while open sourcing DeepSeek Harness, a direct rival to Claude Code: AI money keeps compounding even as Ramp's data shows adoption growth from OpenAI and Anthropic both slowing as spend shifts to open-source models. The Trump administration also escalated its fight to keep Kalshi operating in New York, and Anthropic said it will start invisibly watermarking everything Claude touches to comply with the EU AI Act.

AI / LLMs

• Anthropic's investors expect the AI lab to float at a $2 trillion valuation or more in an October IPO, eclipsing SpaceX as the largest ever: backers point to annualized revenue they expect to reach $100B-$120B by year end, over 10x where it stood in January, though the figure comes from investors rather than an Anthropic target, and the company still faces a Pentagon supply-chain-risk dispute and June's brief export-driven pullback of Fable 5 and Mythos 5.
• DeepSeek open sourced DeepSeek Harness, the MIT-licensed agent framework built on its "everything is a plugin" Cordis meta-framework that it used to benchmark V4 internally, positioning it as a direct, developer-preview rival to Claude Code, the same day it raised API prices ahead of a potential IPO, with V4-Flash output tokens jumping from $0.28/1M to as much as $1.32/1M at peak under dynamic pricing it had never used before.
• Anthropic is in talks to buy Decart AI, a real-time generative video and GPU-optimization startup, for about $6B: folding video generation and inference-efficiency work directly into Anthropic's stack rather than building it in-house.
• Anthropic will start invisibly watermarking any content its Claude models process, not just what they generate, to comply with the EU AI Act's machine-readable-marking rule for AI system providers: the law covers any model released after August 2, with a grace period to December 2026 for older ones, and Anthropic is rolling the mark out globally rather than only to EU traffic.
• Sergey Brin has personally urged key Google AI staff to go all in on Gemini in recent months, with some teams shifting from DeepMind to corporate Google: the consolidation continues weeks after Jeff Dean, Sanjay Ghemawat, Oriol Vinyals and Quoc Le left DeepMind for Jeff Dean's own startup and Demis Hassabis moved to chairman.
• Twitch will now let streamers opt out of having their content used to train Amazon's generative AI models, more than two years after an Amazon executive confirmed the company was already doing so: opting out covers streams, VODs, clips, chat and channel text, though captioning and safety tools keep working regardless, and Twitch told TechCrunch it intends to keep training on creators' videos by default.

Devtools & Infra

• Zed launched Delta, a standalone app that replaces git for AI-agent coding sessions with a CRDT-based version control layer, turning agent work into shared, reviewable, real-time multiplayer threads: built for a world where an agent, not a human, is the one committing code.
• Charity Majors, CTO of Honeycomb, told Pragmatic Engineer her skepticism about AI coding broke in November 2025, crediting Claude Code's harness more than the Opus 4.5 model behind it: she compares AI's 2025 impact on the industry to cloud computing's breakout year in 2010, arguing reliability and verification, not code generation, are now the real bottleneck.
• GitHub profiled how AutoGPT, fielding agent-written pull requests from Copilot, Claude and its own tooling across 150+ open PRs, standardized on AGENTS.md files placed beside the code they govern after discovering agents read what's in the directory in front of them, not project wikis or contributor docs.

Security & Privacy

• A supply-chain attack on LiteLLM, the open-source tool used to route AI-driven software development, exposed terabytes of credentials belonging to major companies including Microsoft, Amazon, Cisco, Samsung and Salesforce: security firms CloudSEK and Hudson Rock say the haul includes cloud keys, repository tokens, SSH keys, Kubernetes secrets, package-publishing credentials and AI provider keys.
• Flock will now require US law enforcement to label every license-plate search with a criminal case number, and CEO Garrett Langley admitted the company "got this one wrong", after a Washington Post investigation found officers misusing its reader network for personal and off-book searches: the license-plate-reader firm says it is assuming more responsibility for how police use its tech instead of treating access as unconditional.
• Designers Isaque Seneda and Gabriel Abrucio released ShieldFont, a typeface that renders normally for human readers while serving AI scrapers a subtly altered, nonsensical version of the underlying HTML: a font-level opt-out from unauthorized AI training rather than a legal or robots.txt one.
• ICE is seeking to spend up to $20M equipping agents with electrified gloves that deliver painful shocks, according to a DHS contract notice: the CTG 5 GLOVE, built by Compliant Technologies and marketed as a "Conductive Distraction and De-escalation" device, with deliveries set for March 2027.

Startups & Industry

• A day after invoking CFTC emergency authority to keep Kalshi operating in New York, the Trump administration moved to override New York's gambling law outright to the same end. The prediction market is separately in advanced talks to raise at least $750M at a $40B valuation, nearly double the $22B it was worth in May, even as it referred 32 possible insider traders to the CFTC in the three months through June, evidence regulators reportedly lack the staff to fully act on.
• Databricks closed a $5B round at a $190B valuation, up from $134B just six months ago, with revenue now past a $7B annualized run rate on 80%+ year-over-year growth: the money is earmarked for Unity AI Gateway (routing and cost control across models), Lakebase (serverless Postgres for agent-built software) and Genie (giving AI access to context spread across an enterprise).
• Cisco's Q4 revenue rose 18% year over year to $17.25B, beating estimates, with AI infrastructure orders from hyperscalers worth $4B and a forecast above expectations for fiscal 2027, while Cerebras's Q2 revenue rose 74% to $180M and it raised its annual forecasts too, yet its stock still dropped 14%+ after hours: strong AI-infrastructure demand no longer guarantees a stock pop.

Research

• A new paper argues AI agent safety should be enforced as a runtime contract, not trained into the model: a preventive layer (sandboxes, permission gates, trajectory monitors) blocking dangerous actions, paired with an evidential layer requiring verifiable proof, like test runs and log captures, that good actions actually happened. The authors ground the case in a survey of 52 documented AI-agent safety incidents and an audit of all 28,560 NeurIPS, ICML and ICLR papers from 2023 to 2025, finding an 8-12x publication imbalance favoring training-time safety work over deployment-time enforcement.

Hacker News

DeepSeek's day dominated the front page: V4 Pro's arrival on OpenRouter drew the single biggest crowd (810 points, 320 comments, discussion), and its newly opened DeepSeek Harness, covered above, pulled its own sizable one (358 points, 161 comments). Codex arriving in the ChatGPT desktop app for Linux, in preview, drew a comparable crowd (363 points, 255 comments). Grok 4.6 scoring 61 on the Artificial Analysis Intelligence Index kept yesterday's launch in the conversation (322 points, 349 comments, discussion). Zed's Delta, covered above, pulled its own sizable agentic-coding crowd.

On security and consent, uBlock Origin's apparent decision to stop chasing Facebook's ad-insertion tricks remained the day's most argued-over thread (353 points, 475 comments, discussion), read by many as a concession that the platform's obfuscation now outpaces filter-list maintenance. A related post cataloged scanners spoofing AI-crawler user agents like ClaudeBot to mass-probe sites for vulnerabilities under a trusted name. In Germany, an advocacy group filed a criminal complaint over Meta's AI glasses, feeding the same consent anxieties around AI hardware. Smaller threads remembered AmigaDOS developer Tim King and admired the shadow-mapping tool Shade Map.

Threads

• AI money kept compounding while adoption cracked underneath it: Anthropic's $2T IPO chatter, Databricks' $190B round and DeepSeek raising its own prices ahead of a listing all landed the same day Ramp's data showed enterprise spend growth from OpenAI and Anthropic slowing as businesses shift budget to open-source models, and even Anthropic's own Fable 5 is only 6% of the tokens businesses buy.
• Regulation caught up with AI from two directions at once: Anthropic's EU-mandated Scarlet Letter watermark and Flock's forced case-number logging (after police misuse) both trace back to oversight arriving after adoption, not before it.
• Coding agents got competing new foundations: DeepSeek open sourced its own Claude Code rival the same day Zed's Delta rebuilt version control around agent sessions and GitHub found maintainers standardizing on AGENTS.md files, three answers to who actually governs an agent's work.
• Government leaned on coercive and platform power alike: the administration moved to override New York's gambling law for Kalshi, ICE sought $20M for shock gloves, and Flock admitted its tracking network needs more guardrails.