Titus's Morning Wire

AI Titus News

WEDNESDAY, JULY 8, 2026
ANTHROPIC'S NEW PRIVACY POLICY LETS IT HAND YOUR CLAUDE CHATS TO POLICE WITHOUT A WARRANT, AND WANTS YOUR FACE TO PROVE YOU'RE HUMAN
★ Must-Read / Watch agentic engineering / build-better-agents📚 Learn state of AI coding / useful technique

⚡ AI Frontier · smol.ai

★ Must-ReadOpenAI cleared to release GPT-5.6 Sol, Terra, and Luna worldwide after the Trump administration lifts its restrictions
The GPT-5.6 family had been limited to a small group of trusted partners at the government's request over biology and cybersecurity concerns. That hold is lifted and OpenAI is expanding preview access globally starting this week.
★ Must-ReadxAI completes its rebrand to SpaceXAI, folding Grok fully into SpaceX
The standalone xAI name is gone five months after SpaceX's all-stock acquisition closed. Musk has said the merger exists to build orbital AI data centers, so the corporate structure now matches that ambition.
📚 LearnSam Altman pitches Washington a 5% stake in OpenAI worth roughly $42.6B
Altman's proposal would give the US government a slice of OpenAI's upside through a sovereign-wealth-style vehicle, and he wants Anthropic, Google, and Meta to cede similar stakes. No company has agreed yet.
Meta ships Muse Image, its first Superintelligence Labs model, across Instagram, WhatsApp, and Facebook
The image model turns plain-language prompts into shareable visuals and powers new Instagram Stories effects. Free for basic use, deeper features sit behind Meta's subscription plans.
📚 LearnMeta's Brain2Qwerty v2 turns brain activity into typed text at 61% word accuracy, no surgery required
Up from about 8% for earlier non-invasive approaches, using a brain-signal-reading MEG headset rather than an implant. Meta says accuracy scales with more training data and points to eventual use for people with communication-limiting conditions.
Google brings Gemini Spark, a proactive personal AI agent, to the Mac
Spark can read local files and work across connected apps like Canva, Dropbox, and OpenTable to complete multi-step tasks without being prompted step by step. Beta, US only, Google AI Ultra subscribers 18 and up.

📺 Watch · latest videos

★ Must-WatchSEE CMUX SOLVE Multi-Agent Orchestration (Claude Code and Pi Agent)
A hands-on look at coordinating multiple coding agents against Claude Code and Pi Agent, the kind of orchestration pattern worth stealing if you're running more than one agent at once.
IndyDevDan · 16.9K views · 633 likes
📚 LearnHow I Make Opus Think Like Fable (5 easy steps)
A practical walkthrough for extracting Fable's operating style into a reusable skill for Opus 4.8, including effort-level tuning and routing cheaper tasks to cheaper models.
Nate Herk · 51.9K views · 2K likes
📚 LearnSo AI is 'thinking' for real
Berman's fresh take on what current models are actually doing when they show reasoning traces, useful context for anyone trying to separate real capability from theater.
Matthew Berman · 1.9K views · 135 likes
📚 LearnThe Best AI Coding Setup Isn't the Most Autonomous One (Here's Why)
Cole Medin makes the case that most people get better results staying in the loop at a middle autonomy level rather than chasing full self-driving-style automation.
Cole Medin · 8.1K views · 288 likes
Photographing the stars with Claude
A photographer's real workflow using Claude to nail exposure settings and stitch a Milky Way panorama over Death Valley, a nice change of pace from the usual coding demos.
Claude
📚 LearnNotebookLM's new Short Video Overviews, explained
A quick look at NotebookLM turning uploaded notes and textbooks into short vertical video summaries, handy for visual learners studying dense material.
Google · 14.6K views · 300 likes

🗣 Voices & Blogs

★ Must-ReadImport AI: Fable writes the fastest GPU kernel ever submitted to KernelBench-Mega
Jack Clark's newsletter covers Fable hitting an 18.71x speedup on a kernel benchmark, plus rising AI automation of freelance work and benchmarks showing hour-long computer tasks getting completed end to end.
★ Must-ReadLilian Weng summarizes 35 papers on harness engineering for recursive self-improvement
A research recap arguing that the need to specify goals and context for agents won't disappear even as more capability gets internalized into the core model, with a rundown of the current best harness design patterns.
📚 LearnSimon Willison: sqlite-utils 4.0 ships database schema migrations
The first major version bump since 2020 adds schema migrations, nested transactions, and compound foreign key support, useful groundwork if you're storing agent state or logs in SQLite.
📚 LearnHamel Husain: 'It's Hard to Eval' Is a Product Smell
Husain argues that when an AI product resists evaluation, that's a design flaw, not an inherent limit, and shows how building for user verification makes a product both easier to test and better.
Most enterprises will hand root cause analysis to AI agents within two years
For decades, software engineers have been the eyes and ears of enterprises, combing through voluminous logs, observability tools, consoles, browser… · The New Stack

🌏 The Wire · Drudge / Breitbart

Zuckerberg tells staff Meta's AI agent push 'hasn't really accelerated' months after cutting 8,000 jobs
At a July 2 town hall, Zuckerberg admitted the reorganization wasn't as clean as planned and said the payoff from Meta's AI bets hasn't shown up yet, though he expects it within three to six months.
SpaceX joins the Nasdaq-100 in record time, less than a month after its IPO
The index shortened its usual three-month seasoning period to fast-track large IPOs, and SpaceX, now carrying the newly rebranded SpaceXAI unit, became the fastest company ever admitted.
AI2 Robotics raises $735M at a $2.8B valuation for its wheeled humanoid AlphaBot
The Shenzhen startup builds both the hardware and the vision-language-action models behind it, betting a wheeled base beats a bipedal one on industrial stability and speed.
South Korea commits $880B over the next decade to chips, AI data centers, and humanoid robots
Samsung and SK Hynix lead a plan to double memory output and build AI data-center capacity, alongside a national goal to grow South Korea's share of the global humanoid robot market from 1% to 20%.
Anthropic pushes Fable 5's move to paid usage credits back to July 12 after user pushback
Fable 5 was due to drop off Pro and Max subscriptions this week and require metered billing at $10 per million input tokens and $50 per million output tokens. The free window now runs five more days.
Japan's largest bank rolls ChatGPT Enterprise out to 35,000 employees
MUFG is deploying ChatGPT across Mitsubishi UFJ Bank's workforce and building toward an AI-native digital bank, including letting customers consult ChatGPT about household finances.

📈 Markets

NVDA 196.93 ▲0.7%
MSFT 388.84 ▲0.5%
GOOGL 367.03 ▲0.2%
AMZN 245.98 ▲0.7%
META 615.58 ▲2.5%
AMD 516.11 ▼6.5%
AVGO 370.78 ▼0.8%
PLTR 134.37 ▲1.4%
SPCX 149.47 ▼6.8%
TSLA 402.90 ▼4.0%

The BriefAnthropic's updated privacy policy took effect today for anyone on a Free, Pro, or Max Claude plan: the company can now hand your conversations to police or other government agencies whenever it decides, on its own, in good faith, that disclosure is warranted, without waiting for a subpoena or a judge's order. The same policy lets Anthropic ask for a selfie, a government ID, or a facial geometry scan to verify you are a real person before it keeps letting you use Claude. None of this touches Claude for Work, Team, Enterprise, or API accounts, the change is specific to consumer plans.

Level UpArmin Ronacher found that Claude Opus 4.8 and Sonnet 5 are more likely than older models to send malformed tool calls with invented fields when they hit a schema they don't recognize, likely a side effect of how they were trained inside Claude Code's forgiving harness. If you build agents on these models, spend some time this week hardening your tool-call validation and error handling instead of assuming the model will always call things correctly. Better Models: Worse Tools