The Manifest
Thinking Machines shipped its first open model as physical supply chains took real hits from an Antwerp toxic gas leak and a fresh Brazil tariff, while enterprise AI debates centered on whether deployed 'agents' are actually agents or just chatbots with better marketing.
Get The Manifest in your inbox
One sharp digest of AI × supply chain. Free. No noise.
Today's Top
- 01Thinking Machines amps up its bet against one-size-fits-all AI with its first open model, InklingTechCrunch AI
- 02TSMC says AI demand can run to 2030 – lifting the outlook for air cargoThe Loadstar
- 03155 hospitalised as toxic gas leak forces Antwerp box terminals to closeThe Loadstar
- 04Anthropic, Blackstone bet the next trillion-dollar AI business is implementation, not just modelsTechCrunch AI
- 05Apple Intelligence approved for launch in China with Alibaba's Qwen AITechCrunch AI
Models & Releases
4 storiesEx-OpenAI CTO Murati's Thinking Machines drops Inkling, a 975B parameter model that leads US labs but trails China
Thinking Machines' first from-scratch model ships fully open under Apache 2.0, positioned as a customizable base rather than a benchmark leader. Watch whether enterprises actually fine-tune it or just stick with closed frontier models.
Gemma 4 gets a stealth update that fixes tool calling bugs and truncated responses under the same name
Google quietly patched tool-calling bugs and truncated responses without a version bump. Pin model versions in production if reproducibility matters to your pipeline.
xAI open-sources 'Grok-Build' on GitHub after massive data breach
A coding CLI silently uploaded SSH keys and password files before xAI open-sourced the codebase in damage control. Audit any coding agent's file access before letting it near credentials.
Bonsai 27B is a full open reasoning model that fits on an iPhone
A 27B model compressed under 4GB while keeping 90 percent of original performance signals on-device reasoning is getting real, useful for field ops apps that can't rely on constant connectivity.
Supply Chain & Ops
5 storiesTSMC says AI demand can run to 2030 – lifting the outlook for air cargo
Chipmakers now expect AI-driven demand for servers and data center gear through the decade, which should keep air freight rates elevated well past any normal seasonal peak. Budget accordingly if air is part of your mode mix.
155 hospitalised as toxic gas leak forces Antwerp box terminals to close
A hydrofluoric acid leak on a container ship shut Deurganck dock and hospitalized over 150 people, a stark reminder of dangerous goods risk sitting quietly in every box terminal. Expect knock-on congestion across North European ports.
US tariff on Brazilian exports 'kind of snuck in on us', says Flexport
A 25% tariff on select Brazilian goods takes effect July 22 with exemptions carved out for items the US doesn't produce. Buyers sourcing from Brazil need to map exposure fast, the exemption list is where the real work is.
FCK, here we go again – cyberattack leaves KFC with empty buckets
Ransomware hit a major cold-chain operator and left KFC without stock, the second such incident since 2018. Cold chain and 3PL cybersecurity need the same scrutiny as ERP systems, not less.
FreightWaves: TL, LTL rates to hit new highs in Q3
Truckload rates hit a cycle high last quarter with LTL rate-per-pound reaching an all-time high, and both are expected to climb further in Q3. Lock in capacity commitments now if you haven't already.
Deals & Market
5 storiesAnthropic, Blackstone bet the next trillion-dollar AI business is implementation, not just models
Anthropic, Blackstone, Goldman and others are betting that embedding engineers inside enterprises beats selling software licenses. If it works, expect every major lab to stand up a services arm.
Apple Intelligence approved for launch in China with Alibaba's Qwen AI
Apple finally gets Apple Intelligence into China by partnering with Alibaba's Qwen instead of its own models, a pragmatic call given the regulatory environment. Watch for similar model-swapping playbooks in other restricted markets.
Microsoft is reportedly training salespeople to talk down OpenAI and Anthropic
Microsoft is coaching sales teams to pitch its own models as cheaper and more efficient than the partners it depends on, a sign the OpenAI relationship is cooling into real competition.
Applied Computing wants to give oil and gas operators an AI model for the entire plant
A $20M Series A for a foundation model built specifically for oil, gas and petrochemical plants suggests vertical, physical-world models are attracting real capital, not just chat assistants.
Indian AI coding startup Emergent becomes a unicorn with $130M Series C
A coding agent startup hit unicorn status on a $120M revenue run rate just over a year after launch, evidence that coding agents remain the fastest path to enterprise revenue in this cycle.
Research & Frontier
5 storiesSPINE: Bridging the Cyber-Physical Gap with Agentic AI
Researchers are targeting the calibration gap between a foundation model's decision-making and an actual robot body. This is the unglamorous plumbing that decides whether embodied AI scales past demos.
Self-Improvements in Modern Agentic Systems: A Survey
A survey framing self-improving agents as systems that convert experience into capability gains with minimal human input. Useful primer for evaluating vendor claims about agents that supposedly learn on the job.
Oracle Agent Memory as an Enterprise Memory Substrate for Long-Horizon AI Agents
Enterprise vendors are treating agent memory as its own infrastructure layer, not just retrieval. Expect memory management to become a real line item in agent deployment costs.
GPT-5.6 Sol reportedly disproves a 30-year-old statistics conjecture in 90 minutes after humans couldn't crack it
A statistics professor used GPT-5.6 Sol to disprove a decades-old conjecture in 90 minutes after the prior model failed over 20 hours. It recombined known methods rather than inventing something new, which is the more honest read of what these systems do.
Interventional Grounding Audits: Black-Box Premise-Dependency Tests for LLM Chain-of-Thought via Predicate Substitution
A method to test whether a model's chain-of-thought actually depends on its stated premises, useful for anyone auditing agent reasoning before trusting it with real decisions.
Org & AI Architecture
5 storiesAgentic orchestration: Enterprise AI organizations have a deployment problem, not a platform problem — and most are calling chatbots agents
Across 101 enterprises, most deployed 'agents' are still chatbot wrappers dressed up in new language. If you're evaluating vendors, ask for proof of reliable multi-step execution, not a demo.
Meta employees sue over layoffs they say were driven by discriminatory AI selection systems
Employees are suing Meta over layoff lists allegedly generated by AI systems that disproportionately hit disabled workers and parental leave takers. This case will shape how much liability shifts to the algorithm versus the humans who deployed it.
OpenAI's Codex now encrypts instructions between AI agents, leaving developers blind to internal delegation
OpenAI's Codex now encrypts what its main agent tells its subagents, meaning developers lose visibility into internal task delegation. Worth flagging for any team relying on audit trails for agentic coding tools.
OpenAI is now using AI to attack its own AI, and it's working better than humans ever did
OpenAI's internal red-teaming model finds successful attacks in 84 percent of scenarios versus 13 percent for human testers. Automated adversarial testing is becoming the real security baseline, not a nice-to-have.
Microsoft patches record number of security vulnerabilities, citing its use of AI
570 vulnerabilities patched in one cycle, with Microsoft crediting AI-assisted discovery. That cuts both ways, the same tooling that finds bugs faster will also be used to find them first by attackers.