AI Full-Duplex Voice, Cheap Coding Models, and Inference Shifts
Your daily shot of what's brewing in AI
Fresh Batch
Bold Shots
Today's biggest AI stories, no chaser
OpenAI released GPT-Live and GPT-Live-1 mini on July 8 as full-duplex models that listen and speak simultaneously, replacing Advanced Voice Mode. Paid Go/Plus/Pro users default to the larger model while free users get the mini variant; complex queries delegate to GPT-5.5 in the background. Global rollout is live on iOS, Android, and web with no API access at launch.
Why it matters: The architectural shift from turn-based to full-duplex enables natural interruptions and backchanneling, but ships without day-one API access, slowing enterprise adoption relative to rivals like Gemini Live.
xAI and Cursor released Grok 4.5 on July 8 at $2/$6 per million tokens, undercutting Opus 4.8 and GPT-5.5 pricing while claiming Opus-class performance on coding and agentic tasks. The MoE model was trained on trillions of Cursor developer-agent interaction tokens and is available in Cursor, Grok Build, and the SpaceXAI console. EU access is delayed until mid-July.
Why it matters: Token efficiency claims (4.2x fewer output tokens than Opus on SWE-Bench Pro) plus aggressive pricing make cost-per-task the new battleground rather than raw benchmark crowns.
Microsoft began routing tens of thousands of Excel and Outlook prompts weekly to its own MAI models, targeting high-volume commodity tasks while reserving OpenAI and Anthropic models for harder work. MAI-Code-1-Flash is rolling out in VS Code and GitHub Copilot; seven new MAI models were shown at Build 2026. The move explicitly aims to reduce and eliminate Anthropic spend.
Why it matters: Commodity inference that scales third-party bills is the first target; securing IP rights through 2032 gives Microsoft a structural off-ramp from future pricing pressure.
The Blend
Connecting the dots across sources
Inference economics now dictate model choice over raw capability
- Grok 4.5 and SWE-1.7 undercutting Opus 4.8 and GPT-5.5 on cost-per-task via token efficiency and per-run pricing
- Microsoft routing commodity Excel/Outlook traffic to MAI models explicitly to cut Anthropic and OpenAI spend
- SambaNova's $11B valuation and JPMorgan on-prem win framed around inference infrastructure economics rather than training
Slow Drip
One standout practical piece per category from this week's posts.
Architectural forks and decisions behind the three leading models.
Mental model and 12 ready-to-run loop recipes from the Claude Code team.
The Grind
One paper per category from current top candidates.
Layer-wise RL study shows single-layer updates can match full-parameter training gains.
On Tap
Skills for agent self-improvement and browser automation are seeing rapid adoption alongside practical inference talks.
Captures learnings, errors, and corrections to enable continuous improvement for the agent.
Fast browser automation CLI for AI agents. Chrome/Chromium via CDP with accessibility-tree snapshots and compact @eN element refs.
3.7K viewsThis video details how Anthropic's Claude Fable 5 autonomously optimized the llama.cpp codebase, achieving a 64% speed improvement on consumer GPUs without hardware changes.
Roast Calendar
Upcoming Bay Area and virtual events focused on agents, hardware, and on-device work.
Robotics and embodied AI builders meetup with strong attendance.
Hands-on reinforcement fine-tuning session for reliable tool use.