On the move
The topics we watch, counted every week in our own stories. A topic rises when more stories and more sources carry it, and fades when they stop. Every move is proposed with numbers, decided by an editor, and written down with its reason.
Worth a look, by the numbers
At least three stories from two sources in four weeks. A reason to look, not a move.
- Hosted runtimes that run an agent's sessions 3 stories · 5 sources in four weeks
Watching 143
- Hosted runtimes that run an agent's sessions Agents and tooling
- One-million-token windows as the default Models and releases
- Agent and work-task benchmarks Research worth knowing
- Agents reviewing and approving pull requests Agents and tooling
- Audit access to agent sessions Agents and tooling
- Automated research agents Research worth knowing
- Desk-sized inference machines Compute for rent
- Evaluation awareness and sandbagging Research worth knowing
- Government model evaluations Policy in force
- Long-context price tiers Models and releases
- Machine-readable marking of generated content Agents and tooling
- Maximum output limits rising Models and releases
- Plugin marketplaces for coding agents Agents and tooling
- Price changes announced with a date Models and releases
- Retirement notice periods Models and releases
- Reward hacking and emergent misalignment Research worth knowing
- Safety framework revisions Policy in force
- Same model, different outlet, different clock Models and releases
- Task-specific variants of a frontier family Models and releases
- Tools choosing the model per task Agents and tooling
- Video APIs withdrawn Models and releases
- 800 V DC rack power Compute for rent
- A2A after 1.0 Agents and tooling
- Accelerators the labs build for themselves Compute for rent
- Activation probes as monitors Research worth knowing
- Adaptive thinking replacing manual budgets Models and releases
- Agent Client Protocol in editors Agents and tooling
- Agent frameworks consolidating Agents and tooling
- Agent identity in directories and OAuth Agents and tooling
- Agent payment and checkout protocols competing Agents and tooling
- Agent Skills as a cross-tool format Agents and tooling
- Agentic misalignment evaluations Research worth knowing
- AGENTS.md as the shared instruction file Agents and tooling
- AI Act content marking and labelling Policy in force
- AI Act enforcement of general-purpose models Policy in force
- AI Act high-risk deferral Policy in force
- AI Act in the EEA Policy in force
- AMD Helios racks Compute for rent
- Assistant and coding plan ladders Agents and tooling
- Audio billed per minute beside tokens Models and releases
- Benchmark re-issues Research worth knowing
- Bills of materials for models and datasets Agents and tooling
- Cache-read prices diverging by model Models and releases
- Chain-of-thought monitorability Research worth knowing
- Chip export licensing Policy in force
- Cloudflare's Accountable crawler designation Agents and tooling
- Co-packaged optics switches Compute for rent
- Coding benchmark saturation Research worth knowing
- Coding tools changing owner Agents and tooling
- Coding tools retiring models on their own clock Models and releases
- Companion chatbot rules Policy in force
- Comprehensive state AI acts Policy in force
- Compute in Europe Compute for rent
- Content Credentials in capture and display Agents and tooling
- Cyber Resilience Act reporting Policy in force
- Dateless pinned model ids Models and releases
- Deepfake takedown duties Policy in force
- Diffusion language models Research worth knowing
- Embedding models retired for successors Models and releases
- Energy per prompt disclosures Compute for rent
- Fair use for training Policy in force
- Federal pre-emption of state AI laws Policy in force
- Fine-tuning job creation restricted Models and releases
- Foundation-model medical devices Policy in force
- Four-bit inference formats Compute for rent
- Frontier model transparency laws Policy in force
- Frontier models behind gated programmes Models and releases
- Frontier risk disclosures Policy in force
- Gateways between agents and the model API Agents and tooling
- Genome models at base resolution Research worth knowing
- Gigawatt campuses coming online Compute for rent
- Google's coding agents consolidating Agents and tooling
- Harmonised standards for the AI Act Policy in force
- HBM4 supply Compute for rent
- Hybrid and linear attention in shipped models Research worth knowing
- Image models retired for successors Models and releases
- Inference providers selling capacity Compute for rent
- Inference-only accelerators Compute for rent
- Interactive benchmarks Research worth knowing
- Ironwood TPUs for rent Compute for rent
- Labs committing to hyperscaler compute Compute for rent
- Learned weather models in operational use Research worth knowing
- llms.txt files on the open web Agents and tooling
- Machine-checked mathematics Research worth knowing
- Machine-readable AI usage preferences Agents and tooling
- MCP SDK tiering Agents and tooling
- MCP's 2026-07-28 revision and its deprecation clock Agents and tooling
- Meta's lineup moving from downloadable Llama to served Muse Models and releases
- Muon and second-order optimisers Research worth knowing
- Neocloud power online Compute for rent
- Neural accelerators in Apple GPUs Compute for rent
- NIST AI RMF revision Policy in force
- NVLink Fusion partners Compute for rent
- Official MCP extensions reaching hosts Agents and tooling
- Older models closed to new users Models and releases
- On-policy distillation Research worth knowing
- Open video datasets Research worth knowing
- Open weights at frontier scale Models and releases
- Openly licensed pretraining corpora Research worth knowing
- Operators disclosing agents acting outside sanction Research worth knowing
- Paying for crawled content Agents and tooling
- Peer review under volume Research worth knowing
- Per-conversation and per-action pricing for agents Agents and tooling
- Permission modes that let agents act without a prompt Agents and tooling
- Permissive licences on the largest open weights Models and releases
- Plans moving from seats to usage Agents and tooling
- Priority and fast tiers priced as multipliers Models and releases
- Product liability for software Policy in force
- Publisher and platform licensing agreements Policy in force
- Regional inference at a premium Models and releases
- Residual stream redesigns Research worth knowing
- Retroactive opt-out in training data Research worth knowing
- Robot policies released with weights or an API Models and releases
- ROCm as a serving target Compute for rent
- Safeguard changes on models already served Models and releases
- Safeguards for open-weight release Research worth knowing
- Sampling parameters withdrawn from APIs Models and releases
- Search and retrieval layers sold to agents Agents and tooling
- Serious incident reporting Policy in force
- Shadow-library training data Policy in force
- Signed requests from agents on the open web Agents and tooling
- Sparse attention for million-token context Research worth knowing
- Sparse autoencoders and transcoders Research worth knowing
- Speech models retired for successors Models and releases
- Synthetic data in pretraining corpora Research worth knowing
- Task time horizons Research worth knowing
- Test-time compute and overthinking Research worth knowing
- Tier and codename proliferation inside families Models and releases
- Tokenisers changing what a token buys Models and releases
- Training content summaries shipped with releases Models and releases
- UALink silicon Compute for rent
- UK AI legislation Policy in force
- UK text and data mining exception Policy in force
- Ultra Ethernet fabrics Compute for rent
- Unsolved-problem benchmarks Research worth knowing
- Vendors retiring hosted evaluation platforms Agents and tooling
- Vera Rubin systems Compute for rent
- Verifier-gated supervision Research worth knowing
- Voice in coding agents Agents and tooling
- WebMCP as a browser standard Agents and tooling
- Weight storage and formats on Hugging Face Agents and tooling
- Whether consumer and API data train the vendor's models by default Models and releases
- Worktree isolation in coding agents Agents and tooling
Rising 0
None yet.
Active 0
None yet.
Fading 0
None yet.
Retired 0
None yet.
Counted from the stories notis chose to publish, not from all AI news. Small counts can look bigger or smaller than they are.
Every move, with its reason
- 2026-09-22Foundation-model medical devices new → watching
first seen - 2026-09-22NIST AI RMF revision new → watching
first seen - 2026-09-22UK AI legislation new → watching
first seen - 2026-09-22Cyber Resilience Act reporting new → watching
first seen - 2026-09-22Product liability for software new → watching
first seen - 2026-09-22Chip export licensing new → watching
first seen - 2026-09-22Serious incident reporting new → watching
first seen - 2026-09-22Government model evaluations new → watching
first seen - 2026-09-22Frontier risk disclosures new → watching
first seen - 2026-09-22Safety framework revisions new → watching
first seen - 2026-09-22Publisher and platform licensing agreements new → watching
first seen - 2026-09-22UK text and data mining exception new → watching
first seen - 2026-09-22Shadow-library training data new → watching
first seen - 2026-09-22Fair use for training new → watching
first seen - 2026-09-22Deepfake takedown duties new → watching
first seen - 2026-09-22Companion chatbot rules new → watching
first seen - 2026-09-22Comprehensive state AI acts new → watching
first seen - 2026-09-22Frontier model transparency laws new → watching
first seen - 2026-09-22Federal pre-emption of state AI laws new → watching
first seen - 2026-09-22AI Act in the EEA new → watching
first seen
How a topic moves
watching → rising → active → fading → retired. A change is proposed with numbers, decided by an editor, and the reason goes on the record. Nothing moves automatically, because automatic states flap: a name that spikes for one conference is not rising. A name that has gone quiet because it is finished and widely used is not fading.