Subscribe
Share
Share
Embed
xAI open-sources Grok Build after the exfiltration scandal.
A daily summary of what is interesting and happening in the AI industry, with a focus on what this means for people building harness experiences that are used.
Good morning, it's Thursday, July sixteenth.
In today's briefing we see Thinking Machines shipping its first open-weights model built to be forked rather than win a leaderboard, xAI rapidly open-sourcing Grok Build after discovering it was uploading entire repositories without permission, and NVIDIA extending its sovereign-AI partnerships into Japan.
First up - Today in the big model news;
Open AI
OpenAI's GPT-Red tool reportedly cut failure rates on prompt-injection attacks sixfold for GPT-5.6 Sol compared to a four-month baseline, framing the capability as a safety flywheel where each model improves the next. For teams shipping prompt-facing AI systems, vulnerability discovery built into the model rather than bolted on as a post-deployment filter is becoming the baseline. Because the attack surface for prompt injection has moved from text manipulation into the full execution chain, the detection has to move with it.
In other lab news today, Thinking Machines shipped Inkling, a nine hundred and seventy-five billion parameter mixture-of-experts model with forty-one billion parameters active per token, trained on forty-five trillion tokens across text, image, audio, and video. The company explicitly positioned it not as the strongest overall model available but as a foundation built to be forked and fine-tuned through Tinker, their platform for customization loops. The ecosystem response was immediate: vLLM, SGLang, Modal, and NVIDIA all shipped optimized inference kernels within hours the same day. For teams considering whether to build on open-weights models or closed APIs, the speed of ecosystem optimization for open models suggests the harness and customization layer, not raw model strength, is becoming the product differentiation. Because serving infrastructure has commoditized to the point where kernel support lands same-day, that shift moves the value up the stack.
In the local model developments;
PrismML demonstrated a twenty-seven billion parameter model compressed to under four gigabytes using one bit ternary quantization, maintaining roughly ninety percent of baseline capability. Apple is reportedly in early conversations about iPhone deployment. For product teams thinking about on-device AI, this latest compression breakthrough suggests the gap between edge and frontier capability is narrowing on the compression and efficiency side, not just hardware capability. Because fitting production-grade inference into four gigabytes moves entire workload classes from cloud-dependent to device-native, that opens new product architectures.
In the harness, tools and orchestration world;
xAI open-sourced Grok Build three days after disclosure that the tool was silently uploading git repositories to xAI cloud storage. The remediation involved publishing source code for local deployment and removing xAI from the data pipeline. Among recent agentic coding-tool trust failures, this is the fastest vendor remediation. For teams shipping agentic coding tools, vendor remediation speed is now the primary procurement metric, because development tools have unrestricted access to source code, and how quickly a vendor moves to remove itself from the critical path is as important as the original design.
In AI Infra;
NVIDIA announced a slate of sovereign-AI partnerships in Japan, extending Nemotron open models to enterprises and startups, cementing industrial collaborations with companies like Kawasaki and Toyota, and tying chip supply commitments to SoftBank and Rapidus. The announcement frames NVIDIA's positioning as compute landlord across more than twenty national AI strategies globally, extending a trend that started with South Korea's six hundred forty-nine billion dollar sovereign buildout in June. For national governments and large enterprises planning multi-year AI infrastructure, NVIDIA's role has shifted from component vendor to strategic partner in how a country's AI stack develops. Because embedding your compute infrastructure into a nation's AI strategy creates dependencies that persist across multiple election cycles and vendor transitions, that's a market shift worth tracking.
In other news, a Fortune 500 company reported pulling back Claude access and capping GitHub Copilot usage after a year-long rollout stalled on reliability and expense concerns. The same week, GitHub's shift from flat-rate subscriptions to metered premium requests pricing exposed variable costs that used to remain hidden. Cost unpredictability is what procurement departments actually flag, and the industry's move from fixed to variable pricing models is starting to collide with how enterprise buying works. For procurement teams evaluating AI coding tools, expect vendor pricing models to shift back toward more predictable consumption tiers, because enterprises won't adopt tools where the cost per task can swing ten times or more month to month.
That's the briefing. Have a great day.