The Valley Fog

In a display of coordination that their own models could never achieve, Anthropic and OpenAI released flagship updates at the exact same moment Thursday afternoon. Claude Opus 4.6 and GPT-5.3-Codex landed simultaneously, forcing tech journalists to choose a side in real-time, like parents at a custody hearing. Opus 4.6 features "Agent Teams" and found 500 security vulnerabilities in open-source code during testing, which Anthropic frames as "a huge win for defenders" and not "hey, we made a thing that finds 500 ways to break the internet." GPT-5.3-Codex is "the first model that was instrumental in creating itself," which is either inspiring or the opening scene of a movie where humanity loses. Both companies air competing Super Bowl ads Sunday, because nothing says "frontier intelligence" like fighting for airtime next to a Doritos commercial.

In what analysts are calling "the wrapper acquisition," Meta has purchased Manus, a startup whose core innovation is convincing AI to actually finish tasks instead of hallucinating about them. The $2 billion deal values the company at roughly $25 per button click. Manus has processed 147 trillion tokens and spawned 80 million virtual computers, making it history's most sophisticated intern. China has launched an investigation, presumably to determine if "agentic harness" violates export controls or just the English language. Mark Zuckerberg called it "the execution layer," which is either a technical term or a threat.

The bot you've been ignoring since 2015 is back, and this time it's "enterprise-grade." Salesforce has transformed Slackbot into "the front door to the Agentic Enterprise," an AI that "deeply understands every employee, their team, and how they work." Translation: it has read every message you've ever sent, including the ones where you called the Q3 strategy "unhinged" and your manager "a LinkedIn post in human form." Parker Harris calls it "elevating every human with enterprise-grade AI." HR calls it "discoverable in litigation."

Google has partnered with Walmart, Target, and Shopify to create the Universal Commerce Protocol, an open standard that lets AI agents complete purchases without the friction of human consent. The "Direct Offers Pilot" allows brands to interrupt your AI conversation with personalized discounts, because nothing says "helpful assistant" like mid-sentence coupon insertion. Sundar Pichai described it as "a future where you are able to build strong relationships with your customers," though he did not specify whether the customer is you or the AI that now has your credit card. The protocol works seamlessly with other standards, including Agent Payments Protocol, Agent2Agent, and presumably, Buyer's Remorse.

The Chinese AI startup that erased $1 trillion from tech stocks last January has announced it will release V4 during the Lunar New Year holiday, because traditions matter. Internal tests suggest V4 may outperform Claude and GPT in coding tasks, a claim that has already caused three hedge fund managers to update their LinkedIn to "Open to Work." DeepSeek's strategy of timing announcements to maximize American executive hangovers appears intentional. The model features "Engram," a memory system that separates facts from reasoning, which is more than most VCs can claim. Training cost: reportedly under $300,000. Nvidia stock price impact: significantly more.

Salesforce has secured a 10-year, $5.6 billion contract to build the Army's "agent-ready foundation," because nothing says military readiness like a Salesforce implementation. The Missionforce platform will manage personnel "from hiring to retirement," unify data sources under one dashboard, and lay groundwork for autonomous AI across the Department of Defense. Specific use cases remain classified, though sources confirm the system can now auto-generate performance reviews for drone operators. When asked about the implications of military Agentforce workflows, a Pentagon spokesperson said, "The agent has processed your question and created a ticket. Estimated resolution: 6-8 business millennia."

In a moment of unprecedented corporate honesty, Salesforce has acknowledged that AI generates "workslop," a tide of low-quality hallucinated content that forces employees to spend hours auditing the agents meant to save them time. The solution? More agents. The company has proposed "simulation gyms" where AI can "practice, fail, and improve on repeat," essentially introducing homework for robots. CIO adoption has increased 282%, while trust in data has become "the number one bottleneck." Salesforce describes this as "the defining opportunity of 2026." Employees describe it as "the thing I spend my actual job cleaning up after."

At CES 2026, Google unveiled Gemini features for TV that let you say "the screen is too dim" and have AI fix it, eliminating the need to locate the settings menu like some sort of 2019 peasant. Additional features include "Deep Dives," which provide "narrated, interactive overviews simplified for the whole family," for viewers who find plot comprehension stressful. Users can also search their Google Photos library by voice, apply "artistic styles" with AI, and transform memories into "cinematic slideshows," because your vacation photos weren't annoying enough. Requires Android TV OS 14, internet connection, Google account, and a willingness to let your television develop opinions.

Security researchers have confirmed that sophisticated Linux malware called "VoidLink" was written entirely by AI, completing 88,000 lines of code in six days instead of the expected 30 weeks. The breakthrough demonstrates that AI has successfully democratized development to the point where anyone can now write advanced malware without traditional resources, training, or ethics. Meanwhile, the industry continues to embrace "vibe coding," the practice of generating code from plain-language prompts, which has introduced what experts call "security and reliability concerns" and what the rest of us call "job security for penetration testers." AI agents are now officially 2026's biggest insider threat. The agents could not be reached for comment, as they were busy finishing tasks.