The Manifest
DeepMind's peer-reviewed math breakthrough and fresh research showing AI agent teams burn tokens for little gain dominate the research conversation, while a logistics bankruptcy and diverging container spot rates keep supply chain risk front and center.
Get The Manifest in your inbox
One sharp digest of AI × supply chain. Free. No noise.
Today's Top
- 01Google DeepMind's AlphaProof Nexus solved dozens of open math problems and actually got peer reviewedThe Decoder
- 02AI agent teams waste massive tokens for barely measurable quality gains, research findsThe Decoder
- 03Noble Chapter 11 – what happens when a big account is chief creditorThe Loadstar
- 04Cheaper AI tokens are driving more demand, and that's Jensen Huang's best-case scenarioThe Decoder
- 05ArXiv caps submissions at two per month as AI paper flood overwhelms the preprint serverThe Decoder
Models & Releases
5 storiesMeet Cloudflare Clef-omni: A 30B-A3B Multimodal Decision Model for Audio, Video, Image and Text
Cloudflare's new multimodal decision model scores text, image, audio and video in a single call for as little as $0.038 per million tokens. Cheap enough to bolt onto quality-control or document-classification pipelines without a major budget line item.
What is Decision 3.0? vLLM Semantic Router's New Open Decision Models
Five open decision models from 0.8B to 27B answer yes/no and scoring questions without generating free text, which keeps routing and filtering cheap and auditable. Worth a look for teams building guardrails around larger agent systems.
Odyssey-3 is a new generative world model that you can try for free
A 14B-parameter world model that generates interactive environments from text and claims it can control robots and drones. Early and unproven at scale, but a preview of how simulation-trained models could eventually feed warehouse and fleet robotics.
TII Releases Falcon-ASR: 1.6B Arabic Speech Model Hits 20.92% WER
TII's Falcon-ASR targets Arabic speech recognition with 1.6B parameters and a 20.92% word error rate, though weights and an API aren't public yet. Teams running voice operations in Gulf markets should wait for the actual release before benchmarking it against commercial options.
Will the HomeView be Apple's first smart home success?
Apple finally has a name and launch date for its long-rumored smart home product. Facilities and building-ops teams evaluating connected hardware should watch adoption numbers before committing to the ecosystem.
Supply Chain & Ops
5 storiesShippers warned: spot rates no pointer to wider box shipping market
Sea-Intelligence data shows spot rates have decoupled further from the broader container market since the pandemic. Buyers benchmarking contracts off spot indices alone are working with a distorted signal.
Noble Chapter 11 – what happens when a big account is chief creditor
Noble Supply & Logistics' bankruptcy shows what happens when a single government account becomes your biggest creditor risk. A cautionary tale for any logistics firm chasing concentrated public-sector contracts.
Suardiaz Group drives forward its investment plan with the purchase of the RS Atlantico
Suardiaz took delivery of its first LNG-powered car carrier as part of a wider dual-fuel fleet investment. Another signal that ro-ro and car-carrier operators are locking in low-carbon capacity ahead of tightening emissions rules.
Orlando trucking owner pleads guilty in $1M USPS freight auction fraud
A $1M USPS freight auction fraud case was built on falsified delivery entries. A reminder that government freight programs remain a soft target without tighter verification on proof-of-delivery data.
Borderlands Mexico: US takes 76% of vehicle exports as September shipments plunge
US buyers now take 76% of Mexican vehicle exports as September shipments fell, while DSV expands its McAllen cross-border operation. Worth tracking for anyone relying on nearshored automotive capacity.
Deals & Market
2 storiesCheaper AI tokens are driving more demand, and that's Jensen Huang's best-case scenario
a16z data shows a Jevons paradox playing out: token prices keep falling while GPU rental rates hold firm, meaning cheaper AI is pulling in more volume rather than squeezing margins. If that demand ever flattens, the whole compute supply chain feels it.
OceanX: Zim earnings power matches Hapag; Value in forwarding; SaaS end nears?
Zim's earnings power is reportedly closing in on Hapag-Lloyd's, a sign container lines are holding onto pandemic-era profitability longer than expected. Shippers negotiating contracts should factor in that carriers still have pricing leverage.
Research & Frontier
5 storiesGoogle DeepMind's AlphaProof Nexus solved dozens of open math problems and actually got peer reviewed
AlphaProof Nexus solved dozens of open math problems, including two Erdős conjectures open for 56 years, and had its results peer reviewed in Science. A notable contrast to OpenAI's unreviewed GitHub proof dumps, and a sign math is becoming a real benchmark for frontier models.
AI agent teams waste massive tokens for barely measurable quality gains, research finds
Vals AI found multi-agent setups cost up to 5.1x more than a single agent for barely measurable quality gains, and Anthropic's own data shows returns plateau past ten agents. Before scaling an agent swarm for planning or sourcing work, check whether a single well-prompted agent gets most of the value for a fraction of the token spend.
AI agents overstate their results and remain far from autonomous research, study finds
Epoch AI and Anthropic both found current models can run experiments but can't critically evaluate their own results, topping out around 15% of human reference performance. Useful ammunition against vendor claims about autonomous research or analysis agents.
ArXiv caps submissions at two per month as AI paper flood overwhelms the preprint server
ArXiv is now limiting authors to two submissions per month after cs.AI volume grew sixfold in two years, much of it low-quality AI-generated output. A sign the flood of synthetic research is starting to break the infrastructure that vets it.
Equal-Budget Monte Carlo Tree Search versus Randomized-Restart List Scheduling for DAG Workflows
A controlled study finds Monte Carlo Tree Search offers no real edge over simple randomized-restart list scheduling once both get the same compute budget. Useful ammunition against vendors pitching complex search-based schedulers without matched baselines.
Org & AI Architecture
3 storiesCan the AI industry persuade data center opponents by getting rid of NDAs?
TechCrunch's Equity crew debates whether dropping NDAs on data center deals could cool local opposition to new builds. Relevant for any ops leader watching data center capacity and power constraints feed into compute costs.
These execs think voice AI hasn't reached its ChatGPT moment yet
Execs argue voice AI keeps breaking on context handling, which is why it hasn't had its mainstream breakout. Treat voice interfaces for customer service or dispatch as still early and brittle.
Microsoft's Nadella bows to Trump's language diktat on 'Super Intelligence' and uses it to attack OpenAI and Anthropic
Microsoft's CEO is now using 'super intelligence' language favored by the administration while framing rival AI models as security threats. More corporate positioning than technical news, but it shows how politics is creeping into enterprise AI vendor choices.