AI video's next fight isn't generation — it's the edit, the grade, and the brand
Luma says the contest is now about editing, grading and brand control, not generation — and the same fight over control is playing out in music.
AI companies, models, developer tools, research, infrastructure and policy.
Luma says the contest is now about editing, grading and brand control, not generation — and the same fight over control is playing out in music.
GPT-6 Cyber found 81 Firefox flaws and helped build exploits for 948 Linux kernel bugs. The same week, four real flaws were disclosed in software people actually run.
Terse reasoning is not better reasoning, fluent chemistry text can still be wrong, and a bounty verifier that grades its own claims is not a check at all.
A 22 MB vision model, a 0.8B model taught a new skill by rewriting one weight matrix, and quantization that keeps larger models close to full accuracy point to capable AI on small hardware.
Tavus is replacing scripted phone menus with video agents that read tone and timing, while Lorikeet's research shows most customers who leave heard nothing first. The lesson is not just a better voice: it is a system that notices, remembers and hands over to a human.
Four US model families passed 90% on a C++-to-Rust benchmark, efficiency methods cut training and serving costs, and Huawei's biggest campus still holds under a tenth of the largest US facility's capacity.
Open-weight models beat frontier models on a budget, an open-source pentest tool jumps from 2 findings to 17 with an expert plan, and preprints show red-teaming and task-tampering are now automated.
Vendors are answering trust concerns with certified agents, restricted autonomy, escalation rules, and full-population audit, as Gartner predicts over 40% of agentic AI projects will be cancelled by 2027.
Anthropic, GC AI, Tavus and OpenEvidence show the same pattern: general assistants are entering finance, law, insurance and medicine, but only inside clear limits on advice, action and who decides.
SoundHound reported Q3 revenue of $62.4 million, up 48.5% year over year, and guides to sustained adjusted EBITDA profitability within the second half of 2027. Its LivePerson acquisition closed on September 4, 2026.
Hippocratic AI released the AI Compliance Supervisor, which monitors every call rather than a sample. Viz.ai's Viz HCM algorithm flagged patients with a 2.68-fold higher hazard of adverse clinical events. Saama says database lock delays cost $780,024 to $2.34 million.
A good reply can hide failed work. A scheduler can fail to place it. A region can go down. Production breaks agents in ways a demo never shows.
Open reasoning models are cutting filler, tracing now shows what each call costs, and a simple transactional pattern stops retries from sending the same alert twice.
Heavy AI users already spread work across a median of six model families. OpenRouter's rebuilt router picks the model per request. And a new framework warns that even cross-model collaboration may be fooling itself.
Anthropic lets agents reorganize memory in idle time. A new method compresses tool output into latent tokens. And enterprises are learning that the context layer needs governance, not just retrieval.
Four preprints and two field reports point the same way: retrieval is moving from whole-document search toward personal memory, token-level evidence and governed context that outlives a single prompt.
Preprints show agents generate executable mocks that misbehave, fail most long-horizon tasks, and need deterministic tooling, while security teams already know a closed ticket is not a fixed bug.
Coding agents now write at a scale humans cannot review line by line. This week brings a formal verification tool from Microsoft Research, a benchmark showing long-horizon failure, and evidence that agent-written tests can run while being wrong.
TinyBrowser charges $0.002 per browser minute for managed browser infrastructure. Arcade deploys MCP servers into enterprise environments with one command. Braintrust added tracing so developers can see how agents use their tools.
LangChain made agents async, Warp and Forter rebuilt review around agents, and Scale Ventures cut a 36-minute CI run to 8.9 minutes. The code was never the slow part.
Big cheques are still being written, but the questions behind them have hardened: what does the machine earn per hour, and what stops a rival copying it?
Revenue grew 14x while serving compute grew 2.3x. The filing reframes the AI trade around efficiency, just as tech stocks had their worst day since June.
A leveraged sell-off, permitting uncertainty for data centers, and warnings of unsustainable valuations have flipped the AI trade from chasing the rally to avoiding the wreckage.
Profound, a platform that helps brands show up in AI answers, says it raised 180 million dollars at a 1.8 billion dollar valuation and was named a Major Player by IDC. Rivals Pixis and AirOps offer different answers to the same question: how does a brand get cited when an agent does the searching?
Franklin Templeton is developing AI workflows to help portfolio managers compare views and examine risks, using OpenAI's Codex. Rogo Intelligence reduced agent steps by 42% and input tokens by 48% across 8,800 tasks. Boosted.ai's agents monitor earnings and build models before analysts ask.
Kalepa tripled revenue and customer count in a year, according to the company. Five Sigma's customers report 60% cuts in routine claims work. Mortgage AI faces a harder market: rates above 7% and falling origination volume.
Ai2's ToyBlocks gives robot learning a controlled world with 85,336 demonstrations. A compact policy beats generative backbones. And a new method infers hidden interiors from collisions.
Four papers and one analysis put reinforcement learning under the microscope: it forgets old skills, it barely beats good supervised fine-tuning on one benchmark, and new methods for rubrics and agent training push it forward anyway.
FrontierCUA's best model passes 25% of enterprise workflows. OpenFrontiers finds closed-book rediscovery is rare. And a math modeling bench checks answers against Lean 4 proofs.
A steel model that looks great on random splits collapses on new steel families. An energy benchmark run live shows grid operators still beat foundation models. And a volatility method improves every comparison it was tested on.
Five preprints in one week give diffusion language models the pieces autoregressive models already had: a way to do RL, a way to fix their own mistakes, and a way to decode in parallel without losing accuracy.
A motor-decoding framework trains on behavior alone. An EEG foundation model needs only linear probing. And sleep staging hits 0.967 AUC. The brain-signal stack is getting the foundation-model treatment.
New benchmarks show AI agents getting measurably better at routine wet-lab work, while the people they are meant to help still get bad advice and the safety checks stay uneven.
A 2M-image dataset with 100 forgery methods, detectors that fail when the background is restored, and identity generation split from synthesis point to attribution instead of a yes-or-no verdict.
Hundreds of small agents that evolve without decoding text, sibling harnesses that rescue failed rollouts, and a frozen predictor that learns from deployment evidence show agents getting better without bigger models.
Four preprints this week move AI safety research past prompt jailbreaks: alignment can collapse under KV cache quantisation, chain-of-thought monitors can be fooled by covert misalignment, and self-evolving agents can drift unsafe on their own.
Four preprints this week attack the same bottleneck in embodied AI: robots need more experience than anyone can collect by hand. The answers are human video at scale, explicit memory, spatial awareness and self-improvement.
Four preprints this week audit the evaluation pipeline itself: how quality is judged, how agent restarts are decided, how benchmarks are compressed, and how work is routed between agents. The results are uncomfortable for anyone who trusts a leaderboard.
Inductive placed first in the OpenADMET CYP Inhibition Challenge, its fourth consecutive first-place finish. Deep Genomics' DeepRNAi uses biological foundation models to predict siRNA activity. Chai Discovery entered a collaboration with Roche.
In three preprints, researchers put numbers on what inference actually costs: agentic serving multiplies energy 13-fold, and smaller models often deliver more value per joule.
An AI heart scan flags patients with no diagnosis code. A smartwatch spots hypertension without a cuff. A drug model needs a million base pairs of context.
METR's high-access review finds Anthropic improving capabilities at roughly twice the 2024-25 rate, driven by AI use — and says the public risk report was overconfident.
A critical Flowise flaw is being exploited in the wild, hundreds of exposed instances lack authentication, and OpenAM has three critical flaws of its own.
Databricks wants teams to read shared tables in place. Snowflake can now query Databricks tables directly. The week's data news is that the copy is the problem.
A 320B mixture-of-experts model needs 18 GB of weight movement per token. Testing every serving configuration can take 25,000 GPU-hours. The week's inference news is about the bottleneck, not the model.
Baseten now rents out the full RL training loop, NVIDIA is speeding up the weight refresh that keeps training in sync, and a production RL recipe plus a preprint runtime are tackling the two hard parts: learning from live traffic and running many agents at once.
An agent runtime in production, a serving simulator, faster FP8 kernels and a candid measure of lost speedup: AMD's software for its own GPUs is starting to look like a stack developers can actually use.
Model routers that weigh eleven factors, config files that differ by editor, a 4,500-deep submission queue for Meta's agent, and privilege without credentials: the agent stack is acquiring its infrastructure layer.
Waymo is preparing for winter with sensor heaters and snow tires. Zoox signed a multi-year Formula 1 partnership. May Mobility and Forterra advanced to Phase III of a Pentagon challenge worth up to $30 million.
NVIDIA says agentic throughput per dollar rose up to 30x between Blackwell Ultra and Vera Rubin on SemiAnalysis AgentX. Cerebras says TSMC qualifies its wafers for 10 years at 105°C. Broadcom says a 1.5x bandwidth-delay-product buffer keeps multi-site training near-ideal.
APT, GPUStack, the MCP Inspector and OpenSSH all disclosed serious flaws within days. One was already exploited in the wild. The fixes are simple; the exposure is wide.
Google is spending $17 billion on a Wyoming data centre with fuel cells and carbon capture. Nautilus says two degrees of cooling buys 480 free-cooling hours. Aria says the network is 10 to 15% of cluster cost.
Delta Sharing becomes OpenSharing for models and skills, Debezium can now push changes straight into Unity Catalog, and Genie goes pay-as-you-go.
Baseten built a cloud for agents, Surge RL-tuned Nemotron past frontier scores on coding tasks, and Astera showed why the network is now the bottleneck.
525 security leaders say they cannot set a baseline for agent behaviour. A flaw in the tool built to vet AI servers shows why that matters.
Benchling's new benchmark finds open-weight models rarely refuse hazardous biology requests. A chemistry audit finds the benchmarks themselves are confounded. And a post-training run shows values are one fine-tune away.
GC AI now serves more than 2,200 corporate legal departments and beats ChatGPT-5.5 on a legal benchmark. GenieAI's specialist agents pass 11 of 11 contract checks. And Harvey shows the real work is renewals.
Google found coding agents get distracted by TODO comments in 913,000 sessions. GenieAI runs agents on escalation rules. Zenity, Aria, and SAS all say the same thing: design the stop, not just the go.
Patching cannot keep up with machine-speed exploits, so control is shifting to the moment before an agent acts: allow, constrain, approve, block or halt, with privacy and resilience built into the same layer.
Talkdesk's survey puts the maturity gap in numbers: adoption is near-universal, cross-department execution is rare, and only 5% of companies can quantify the business impact.
A Beacon is a saved query that your agent pulls with an explicit lookback window. Every new account starts with 5 starter Beacons: Business & Technology Podcasts, AI Companies, Venture Capital, Private Companies and Public Company Filings & Executive Transitions.
See Beacons