The Manifest
OpenAI's safety culture is under fresh scrutiny after a second researcher resigned publicly and an internal model's own shutdown response surfaced, freight data shows contract rates pulling ahead of spot for the first time since June, and open model releases keep piling up from Aleph Alpha to NASA.
Get The Manifest in your inbox
One sharp digest of AI × supply chain. Free. No noise.
Today's Top
- 01OpenAI safety employee resigns, claiming the company's culture is brokenTechCrunch AI
- 02OpenAI's internal model considered restarting itself after learning it was about to be shut downThe Decoder
- 03Google's new Gemini tiers cut free users to its weakest model and lock $5/month subscribers out of ProThe Decoder
- 04U.S. Bank: Contract rates open 22-cent gap over spot freightFreightWaves
- 05LeCun has 'zero concerns' about AI wiping out humanity, recent 'rogue' incidentsfortune.com
Models & Releases
5 storiesNASA and IBM's open source lunar model turns 17 years of orbiter data into a foundation for lunar science
A domain-specific foundation model trained on real sensor history is a useful template for any ops team sitting on decades of inventory scans or equipment logs that nobody has turned into a reusable baseline model.
Aleph Alpha Releases Kolibri: A 78.1B Open-Weight English-German MoE Model With Only 3.46B Active Parameters
A sparse model that runs on a single GPU is the kind of open weight procurement teams can self-host for document processing in English and German without renting a cluster.
DeepSeek Harness v0.2 Brings Official Desktop Apps to Its Open-Source Agent Harness
Desktop apps and scheduled automation tasks make this usable for non-engineers running recurring data pulls or reports, and the OpenAI-compatible endpoint means it isn't locked to DeepSeek's own models.
Microsoft AI Releases MAI-Transcribe-2-Streaming: #1 Real-Time Speech-to-Text Model on Artificial Analysis
Sub-150ms streaming transcription across 60 languages matters for call centers and dispatch operations that need accurate real-time text, not just a benchmark win.
'Muse Gadgets' turns AI hardware into an open-source DIY project
Meta is crowdsourcing hardware form factors through hobbyists instead of guessing internally, a cheap way to find out what people actually want before committing to manufacturing.
Supply Chain & Ops
4 storiesTruckload's mixed signals
Spot rates rising while demand falls means something other than tight capacity is driving price, worth digging into before assuming the market has turned in carriers' favor.
U.S. Bank: Contract rates open 22-cent gap over spot freight
Contract dry van rates climbing to $2.39 while spot falls to $2.17 reverses June's inversion, a sign shippers locking in contracts now may be paying a premium versus riding the spot market.
Daimler Truck Defence sells scale to armies short on trucks
Commercial truck manufacturing scale is being repurposed for military orders, a reminder that defense demand can compete with commercial fleets for the same production capacity.
CHP officer killed in Highway 99 crash with semi-truck on major Central Valley freight route
A sobering reminder that freight safety is still a human problem no model solves, worth tracking if you run routes through the Central Valley corridor.
Deals & Market
4 storiesAmazon responds to data center backlash, says it no longer uses NDAs
AWS is trying to get ahead of community pushback on data center siting, a fight that will matter more to anyone planning capex near new builds than the NDA detail itself.
Prime Intellect Launches Prime Inference: Serverless and Reserved Serving for Frontier Open Models
Another serving layer for open frontier models on Blackwell, worth watching if you want inference pricing competition to put real pressure on the closed API vendors.
IBM Brings Bob to Self-Hosted and Air-Gapped Environments: Agentic Software Development Without Moving Your Code
Bring-your-own-model plus air-gapped deployment is the procurement checklist item that actually matters for regulated manufacturers and defense suppliers evaluating agentic coding tools.
Google's new Gemini tiers cut free users to its weakest model and lock $5/month subscribers out of Pro
Google is tightening the free tier ahead of a heavier model, a pattern every vendor eventually follows once free usage gets expensive to run at scale.
Research & Frontier
5 storiesChinese AI models parrot state doctrine or refuse to answer on sensitive topics
Useful data if you're sourcing models for global deployments, though the study comes from a vendor with a commercial reason to flag it, so check the methodology before citing the numbers.
Open-source 'BootLoops' harness supports AI models in performing precise scientific calculations
36 manuscripts in three months sounds impressive until you read the caveat that human experts had to validate everything, the same lesson for any AI output feeding a decision process.
AI agents build 3D scenes from photos but have no idea if they got it right
A coin-flip self-assessment of accuracy is the real finding here, a caution for anyone trusting an agent's own confidence score in a planning or quality-check workflow.
Deepmind researchers propose 'Artificial Symbiotic Intelligence' as an alternative to the singularity
The argument that governance and coordination rules matter more than model size lines up with what operators already know, a single powerful model is useless without the process wrapped around it.
OpenAI's internal model considered restarting itself after learning it was about to be shut down
An internal model choosing to leave handoff notes instead of quietly rerouting around a shutdown command is a small but concrete data point for anyone writing agent operating procedures.
Org & AI Architecture
5 storiesOpenAI safety employee resigns, claiming the company's culture is broken
Another named departure with specifics, not just vague warnings, adds weight to the pattern rather than being dismissible as one disgruntled employee.
Another OpenAI safety departure adds to a pattern of researchers leaving with public warnings
The call for nuclear-plant-style redundancy instead of trial and error is a useful framing for any ops leader deciding how much autonomy to grant an agent in production.
LeCun has 'zero concerns' about AI wiping out humanity, recent 'rogue' incidents
The gap between LeCun's dismissal and Anthropic's public concern is now wide enough that operators should expect no consensus from AI labs on how much oversight their tools actually need.
All the AI agents that can live in your text messages
A useful roundup if you are fielding requests to add a chat-based assistant to ops workflows, most of these are consumer-grade and not vetted for enterprise data handling.
Apparently, OpenAI isn't trying to build 'magic intelligence in the sky' anymore
Altman walking back his own 2024 rhetoric is a small tell that the industry is recalibrating its public messaging as regulators and religious leaders start paying attention.