Titus's Morning Wire

AI Titus News

FRIDAY, JULY 3, 2026
MICROSOFT BETS $2.5 BILLION ON THE HUMANS, NOT THE MODELS: NEW 'FRONTIER COMPANY' EMBEDS 6,000 ENGINEERS INSIDE CUSTOMERS TO ACTUALLY MAKE AI WORK, DAYS AFTER AMAZON COMMITTED $1 BILLION TO THE SAME PLAYBOOK
★ Must-Read / Watch agentic engineering / build-better-agents📚 Learn state of AI coding / useful technique

⚡ AI Frontier · smol.ai

★ Must-ReadAWS puts $1B behind 'forward deployed engineering,' sending pods of engineers into customers
AWS is committing $1 billion to a Forward Deployed Engineering unit that drops small pods of five or six engineers into a customer for 45-day, agentic-first engagements, then leaves the team self-sufficient when it exits. Early customers include the Allen Institute, Cox Automotive, the NBA, the NFL, Ricoh, and Southwest Airlines. It is the first hyperscaler to formalize the embedded-engineer model, and Microsoft matched it two days later.
★ Must-ReadOpenAI is putting GPT-5.6 Sol on Cerebras at up to 750 tokens per second this month
OpenAI's new flagship GPT-5.6 Sol is heading to Cerebras hardware at up to 750 tokens per second in July, part of a staggered Sol, Terra, and Luna rollout that started with about 20 preview partners after review with the US government. General availability is expected in the coming weeks. Frontier reasoning is about to feel a lot less like waiting.
Joulent raises $1.75B to build power infrastructure for AI's compute boom
Houston-based Joulent pulled in $1.75 billion, one of the week's largest rounds, to build energy infrastructure aimed squarely at the power demands of AI data centers. A reminder that the binding constraint on the model race is increasingly electricity, not ideas.
A researcher used Claude to slip past a firewall and expose Ticketmaster's Front Gate ticketing
Security researcher Ian Carroll, working inside Anthropic's Cyber Verification Program, had Claude craft nested-subquery payloads that bypassed a web application firewall and reached 500-plus database tables at Ticketmaster-owned Front Gate Tickets, the platform behind Lollapalooza, Bonnaroo, and SXSW. Front Gate patched it within a day. A clear look at AI-accelerated offensive security operating inside guardrails.

📺 Watch · latest videos

★ Must-WatchSonnet 5 (Fully Tested): It Underperforms GLM-5.2 and Costs More!?
A hands-on test pitting Claude Sonnet 5 against GLM-5.2 on real coding tasks, arguing the open-weight model wins on both quality and price. Spicy, and worth watching before you pick a default.
AICodeKing
★ Must-WatchGLM 5.2 in Claude Code is Blowing My Mind
Running GLM-5.2 as the model inside Claude Code and walking through where the open-weight model holds up against Anthropic's own on agentic tasks.
Nate Herk | AI Automation
📚 LearnI Benchmarked the New Sonnet 5. The Results Shocked Me.
A structured benchmark of Sonnet 5 across coding and reasoning tasks, with the numbers laid out so you can judge the hype for yourself.
How I AI
📚 LearnGLM 5.2 Is INSANE. Better than Claude Fable 5?
A walkthrough testing whether GLM-5.2 can hang with Claude Fable 5 on harder tasks, part of the week's broader open-weight-versus-frontier bake-off.
Zubair Trabzada | AI Workshop

💻 Builder Desk · Hacker News

Virginia bans sale of geolocation data
888 pts · 132 comments
Podman v6.0.0
580 pts · 228 comments
Half-Baked Product
559 pts · 152 comments
Immich 3.0
507 pts · 248 comments
Since Linux 6.9, LUKS suspend stopped wiping disk-encryption keys from memory
505 pts · 215 comments
CarPlay Is Additive
415 pts · 557 comments
An American Privacy Emergency
358 pts · 106 comments

🗣 Voices & Blogs

★ Must-ReadSimon Willison: releasing llm-coding-agent 0.1a0
Willison ships an early coding agent built on his llm tool, a readable, hackable take on what a minimal coding agent actually needs. Good for anyone who wants to understand agents by reading a small one instead of a framework.
★ Must-ReadShashi Jain: AWS kept its deployment engineers in-house while rivals sold equity to get theirs
A sharp analysis of why AWS built its forward-deployed unit internally while competitors leaned on equity-for-talent deals. The clearest lens on what this week's Microsoft and Amazon deployment moves actually mean.
📚 LearnSimon Willison: using DSPy to evaluate and improve an agent's SQL system prompts
A hands-on writeup of using DSPy to measure and then improve the SQL prompts behind the Datasette Agent. A concrete example of treating prompts as something you test and optimize, not hand-tweak.
Semgrep: 'we have Mythos at home,' GLM-5.2 beats Claude in our cyber benchmarks
Semgrep's security team benchmarks GLM-5.2 against Claude on real vulnerability-finding tasks and finds the open-weight model surprisingly strong. A useful, skeptical data point in the open-versus-frontier debate.
The $1.3 million theft that exposed AI’s blind spot
Cyberattacks used to be the biggest security issue surrounding AI infrastructure, but that could be changing. A recent cargo theft The post The $1.3… · The New Stack

🌏 The Wire · Drudge / Breitbart

White House issues an executive action to promote AI innovation and security
A new presidential action pairs a push for faster domestic AI innovation with tighter security requirements, the federal frame behind this year's wave of staggered, government-reviewed model releases. Worth reading for how Washington is trying to have speed and safety at the same time.
North Carolina releases a statewide AI roadmap built around workers and the economy
Governor Josh Stein rolled out an AI Strategic Roadmap aimed at protecting and better serving residents while preparing the state's workforce and economy for AI. Another state treating AI as public infrastructure rather than a vendor pitch.
Physicists make quantum time appear to flow backward in a lab
A new experiment demonstrates a reversible arrow of time at the quantum scale, a genuinely strange result that could inform how we build and error-correct quantum computers.
A spray-on powder can stop life-threatening bleeding in about one second
Researchers developed a powder that halts severe bleeding almost instantly, a cheap and portable idea with obvious value for trauma care in the field.
Old muscle stem cells can act young again, with one catch
Scientists coaxed aged muscle stem cells back toward youthful behavior, a promising thread for treating age-related muscle loss if the tradeoff can be managed.
A strange LIGO signal could point to the missing link behind dark matter
An unusual gravitational-wave signal has revived hopes that primordial black holes are real and might account for some of the universe's dark matter.
Bumble bees just solved a classic animal-intelligence test
Bees passed a cognition test long used to gauge smarts in larger animals, one more dent in the idea that a big brain is required for clever problem-solving.
Scientists find a protein switch that burns fat and blocks new fat cells
A newly identified protein switch both burns existing fat and prevents new fat cells from forming in lab models, a potential target for metabolic disease.

🤖 Trending Models · Hugging Face

deepseek-ai/DeepSeek-V4-Pro-DSpark
text-generation · ★321 · 9.4K 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 BriefThe theme today is that the hard part of AI has shifted from the model to the delivery. Inside of three days Microsoft put $2.5 billion behind a new Frontier Company that embeds 6,000 engineers directly inside customer teams, and Amazon committed $1 billion to a nearly identical forward-deployed engineering unit. Both run any model and get paid on results, meaning they only win when the AI actually moves a business number, not when a contract gets signed. For a working person the signal is plain: the scarce skill is no longer picking the smartest model, it is being the person who can wire one into real work and prove it changed something.

Level UpAnthropic's new Claude Science workbench pulls 60-plus scientific databases and research tools into one place, and it is funding up to 50 projects with as much as $30,000 in credits each, with applications open through July 15. You do not have to be a scientist to steal the idea: the point is turning one-off AI answers into a repeatable, auditable pipeline. Pick one analysis you currently redo by hand each week and try rebuilding it as a workflow you can run again. Claude Science

Caught Up?
  • Microsoft and Amazon both stood up billion-dollar forward-deployed engineering units, moving the AI fight from which model to who can deploy it.
  • California signed a first-of-its-kind deal putting Claude in every state agency at half price; North Carolina followed with its own statewide AI roadmap.
  • Anthropic made Sonnet 5 the default in Claude Code (1M context, $2/$10) and launched Claude Science with up to $30k in project credits.
  • OpenAI began a staggered, government-reviewed rollout of GPT-5.6 Sol, Terra, and Luna, with Sol heading to Cerebras at up to 750 tokens per second.
  • Open-weight GLM-5.2 kept narrowing the gap on frontier coding models, setting off a week of head-to-head tests against Sonnet 5 and Fable 5.