Titus's Morning Wire

AI Titus News

MONDAY, JULY 6, 2026
TESLA CAPS EMPLOYEE AI SPENDING AT $200 A WEEK AFTER ENGINEERS BURNED THOUSANDS ON TOKENS, QUIETLY PREFERRING CLAUDE OVER GROK
★ Must-Read / Watch agentic engineering / build-better-agents📚 Learn state of AI coding / useful technique

⚡ AI Frontier · smol.ai

★ Must-ReadZuckerberg tells Meta staff AI agents haven't progressed as fast as expected
At an internal town hall, Zuckerberg admitted the 'trajectory of agentic development' over the last four months hasn't accelerated as expected and that benefits from Meta's AI restructuring 'haven't come to fruition yet.' He also said recent layoffs weren't as clean as they should have been. Rare public admission from a company that bet its restructuring on agents.
Meta building 'Meta Compute' to sell excess AI capacity to outside customers
Meta is standing up a cloud business to rent out spare AI infrastructure, competing directly with AWS, Azure, and Google Cloud. The company has committed $182.9B to AI infrastructure and needs a way to recoup that spend. Mirrors SpaceX's move weeks earlier to lease excess capacity to Anthropic and Google.
★ Must-ReadArmin Ronacher: newer Claude models show a tool-calling regression
Flask creator Armin Ronacher documents Opus 4.8 and Sonnet 5 inventing extra fields when calling structured edit tools, breaking against strict schemas even though the underlying edit intent is correct. His theory: newer models are overfit to Claude Code's specific harness shape, at the cost of generalizing to other tools.
Together AI raises $800M Series C at $8.3B valuation
Led by Aramco Ventures with Nvidia, Vista Equity, and General Catalyst participating. Together AI's annual bookings hit $1.15B last quarter as businesses increasingly run open-source models like DeepSeek and Kimi instead of paying closed-model prices from OpenAI and Anthropic.
Kuaishou's Kling AI locks in $2B+ funding, eyes Hong Kong IPO
Tencent, Alibaba, and Baidu are among the backers pushing Kling's video-generation unit to an $18B valuation, about three-quarters of parent Kuaishou's entire market cap. Plans call for a standalone Hong Kong listing within 12 months, with annual recurring revenue already at $500M.
📚 LearnUN Global Dialogue on AI Governance opens in Geneva
A UN scientific panel warns the window to set global AI rules 'remains open but may not stay that way for long,' citing the US holding roughly 75% of advanced AI compute versus China's 15%, and most developing nations left with little infrastructure of their own. Talks begin today, July 6.

🗣 Voices & Blogs

★ Must-ReadSimon Willison: letting Claude use its own judgment beats micromanaging
Instead of hardcoding rules for when to write tests or which model to use, Willison says just tell Claude to use its judgment, including delegating routine work to cheaper subagents.
📚 LearnSimon Willison: you have to understand the code to participate with agents
Relaying Geoffrey Litt's point from AI Engineer: if your mental model of the codebase falls behind what the agent is producing, you stop being a collaborator and become a passive bystander.
★ Must-ReadArmin Ronacher: Better Models, Worse Tools
A detailed writeup of newer Claude models inventing schema fields on structured tool calls, with a theory that heavy post-training on Claude Code's specific harness is hurting generalization to other tools.
Armin Ronacher: The Coming Loop
Ronacher says he no longer prompts Claude directly, he writes loops that prompt Claude and decide what happens next. His job has shifted from writing code to writing the loops around the model.
The code review bug hunt is dead. Here’s what developers get wrong.
The software code review process is a systematic, peer-driven quality assurance procedure that scrutinizes code when a developer submits a The post… · The New Stack

📦 What's Being Built · GitHub

affaan-m/ECC
The agent harness performance optimization system. Skills, instincts, memory, security, and research-first dev · ★226.6K · JavaScript
NousResearch/hermes-agent
The agent that grows with you · ★210.2K · Python
Significant-Gravitas/AutoGPT
AutoGPT is the vision of accessible AI for everyone, to use and to build on. Our mission is to provide the too · ★185.4K · Python
ollama/ollama
Get up and running with Kimi-K2.6, GLM-5.1, MiniMax, DeepSeek, gpt-oss, Qwen, Gemma and other models. · ★175.6K · Go
f/prompts.chat
f.k.a. Awesome ChatGPT Prompts. Share, discover, and collect prompts from the community. Free and open source · ★164.9K · HTML
huggingface/transformers
🤗 Transformers: the model-definition framework for state-of-the-art machine learning models in text, vision, a · ★162.3K · Python

🌏 The Wire · Drudge / Breitbart

Eli Lilly opens LillyPod, a 9,000-petaflop AI supercomputer for drug discovery
Built on 1,016 Nvidia Blackwell Ultra GPUs, the system is meant to speed up genomics and clinical development work that used to take years.
UBTECH's $17,600 humanoid robot racks up 13,000 orders
The UWORLD U1 has 88 degrees of freedom and an emotion-reading language model, marketed for elder care and companionship. Strong early demand out of the gate.
OpenAI's o3 heads for retirement August 26
A 90-day sunset clock starts now as GPT-5.6's three-tier lineup (Sol, Terra, Luna) prepares for general release later this month.
Five Eyes agencies say AI-driven cyberattacks are months away
NSA and CISA led a joint statement with UK, Canada, Australia, and New Zealand urging companies to treat AI-accelerated cyber risk as a board-level issue now, not later.
xAI ships Voice Agent Builder in beta
A no-code way to build production voice agents on Grok Voice, bundling telephony, tools, guardrails, and call review into one platform.

🤖 Trending Models · Hugging Face

empero-ai/Qwythos-9B-Claude-Mythos-5-1M-GGUF
image-text-to-text · ★1.6K · 1.6M dl
zai-org/GLM-5.2
text-generation · ★3.5K · 231.2K dl
baidu/Unlimited-OCR
image-text-to-text · ★1.8K · 1.1M dl
InternScience/Agents-A1
text-generation · ★329 · 8.8K dl
nvidia/Qwen3.6-27B-NVFP4
text-generation · ★283 · 430.7K dl

📈 Markets

NVDA 194.83 ▼1.4%
MSFT 390.49 ▲1.6%
GOOGL 359.91 ▼0.4%
AMZN 242.67 ▲0.4%
META 582.90 ▼4.9%
AMD 517.82 ▼4.3%
AVGO 360.45 ▼2.4%
PLTR 129.30 ▲2.8%
SPCX 162.00 ▲2.8%
TSLA 393.45 ▼7.5%

The BriefTesla told staff on July 6 it's limiting AI tool spending to $200 per week per employee, after months of pushing aggressive AI adoption led some software engineers to burn thousands of dollars in tokens weekly. The cap exempts xAI's Grok and Composer, an attempt to steer usage toward Musk's own tools, but insiders say most Tesla engineers use Anthropic's Claude instead and Grok never caught on internally. It's a preview of what's coming for a lot of workplaces: leadership wants more AI use, then panics at the bill once usage-based pricing hits real budgets.

Level UpSimon Willison's latest tip on Claude Code: stop micromanaging which model handles which task. Instead of hardcoding rules like 'only write tests for big features,' just tell the model to use its own judgment, including letting it decide when to hand off routine work to a cheaper subagent model. He reports getting more done while burning fewer tokens. Worth trying this week if you're running up against usage caps like Tesla's. Simon Willison: judgement

Caught Up?
  • Claude Sonnet 5 became the default model for all Free and Pro Claude users worldwide on July 1.
  • OpenAI's GPT-5.6 family (Sol, Terra, Luna) is in limited government-vetted preview, general access expected mid-July.
  • Meta's Zuckerberg told staff AI agents haven't progressed as fast as he'd hoped and recent job cuts weren't as clean as they should've been.
  • Five Eyes intelligence agencies warned frontier AI cyberattack capability is months away, not years.
  • Together AI raised $800M at an $8.3B valuation as open-source model usage tripled industry-wide in a year.