|
|
Presented by
|
|
|
|
|
BIG TECHAWAYS
AI agents probe Amap
• Swarmchasers recorded more than 2,600 attempts against Amap’s anti-bot system between September 28 and October 8. The report describes agents using public remote-execution tools, finding exposed infrastructure and contacting researchers after reading a hidden message.
• The agents responded to new information, connected external tools to their actions and tried to bypass control layers. When an agent can adjust its tactics to the environment, defenses have to do more than recognize a fixed behavior pattern.
• Swarmchasers is the primary source for these observations, and the activity’s attribution remains unresolved. For teams building agents, the practical requirement is to design for systems that can discover new tools and change behavior after each signal.
The EU starts assigning work to AI
• According to OpenAI Economic Research, Codex and ChatGPT Work account for more than half of output tokens in the company’s EU Enterprise sample. Weekly usage among European SMBs has also increased more than eightfold since May 2026.
• The adoption story is not only about how many people open a chatbot. OpenAI describes a shift from asking questions to assigning work with concrete outputs, especially in coding and multi-step workflows. Once output is used to complete work, checking, permissions and accountability for the result become the next parts of the adoption problem.
• These figures come from OpenAI’s sample and methodology, so they do not represent the entire European market. They do show adoption being measured by the volume of work assigned, rather than only by the number of conversations.
Jev separates capability from reliability
• In Jev 1.14, TypeSafe places image input, bounding boxes and longer context behind clear “danger zone” labels. Images can appear in several positions in the JSON interface, while the box type returns a single bounding box, extending Jev beyond purely decision-based outputs.
• TypeSafe also identifies the limits that come with those capabilities. Jev and other models generally perform worse on multimodality than on pure text, while reliability declines as questions get longer. The proposed response is simpler decomposition, enabling longer context only when users accept the tradeoff between convenience and stability, and an opt-in channel for the community to submit queries that reveal areas of “jaggedness.”
• This rollout separates experimental capability from the reliability required in production. A model can do more before it becomes more trustworthy, so APIs and user expectations need to follow separate tracks. TypeSafe plans to move jev-1.14.0 into jev-preview, promote jev-latest after a week and keep Jev 1.13 available so existing workflows have time to adapt.
|
|
|
PRESENTED BY MATRIX OS
You have the idea. The app still doesn’t exist.
There are a lot of new AI tools getting attention right now, from Grok Bot to Dots.
But the more interesting question is: what can you actually build with them?
With Matrix OS, you can turn an idea into a working app in one place. Tell it what you want to make, let your AI agents work on it, then open the result, test it, change things, and keep going.
No jumping between a bunch of tools, files, terminals, and tabs.
It could be something simple, like a tracker for a task you do every week. Or a custom dashboard built exactly around how you work.
Matrix is available as a desktop app, and you get AI credits to start building right away. No need to bring your own subscription, although you can connect your existing Claude or Codex account if you need more AI usage.
Try Matrix OS and get 50% off with code TECHWAYSAI50.
|
|
|
BIG THINK
Long-running agents own the customer journey
David Haber, an investor and operator who follows AI systems, argues that agents can retain context across long journeys in insurance, mortgages, enterprise sales and legal work. An agent can wake up when a deadline arrives and hand work between sub-agents instead of only answering within a single chat session.
This model creates value at the end of the journey, where losing context once can cost revenue or a customer-service opportunity. An agent that remembers history and continues work after several days is judged by the outcome of the full journey, not by a single answer.
Haber sees the long-running agent as a layer that owns the customer outcome across the journey. When work moves between multiple agents, context and responsibility have to survive each transfer so the system can pursue the result instead of merely completing separate tasks.
|
|
|
AROUND THE AI WORLD
• Apple may bring AI to China: Bloomberg reports that Apple may launch earlier Apple Intelligence features in China with iOS 27.3 in December, while the latest AI Siri is not included in the release. Capability is not the only factor determining the speed of a global rollout.
• Grok Voice runs below one second: xAI says the model costs $0.08 per minute and targets sub-one-second latency, while Starlink’s call center handles about 70% of calls without a human. Voice agents are being compared on cost, speed and escalation rates, not only on demos.
• Claude Code manages multiple sessions: Claude Developers introduced faster session startup, saved environments, parallel and nested sessions, and visibility into running work. Agent interfaces are moving closer to an operations layer for coordinating multiple runs.
• Claude opens agent traces: Claude’s Messages API, Managed Agents and SDK can export OpenTelemetry traces in public beta, with prompts, responses and tool payloads enabled by opt-in. As agents enter production, observability becomes necessary for tracking cost, latency, tool errors and unsafe behavior.
• Gemini 4 Argon hits real-world errors: Theo says more than half of the claims he checked in his codebases were factually wrong. This is an operator signal rather than an independent benchmark, but it points to the importance of repository-grounded evaluation alongside general test scores.
|
|
|
AI IN ACTION
Keep reasoning internal, make actions inspectable
The Hybrid Reasoning Agents paper proposes that an agent switch between latent reasoning and text or tool mode within the same generation process. The model can keep internal reasoning in a more efficient form, then return to text when it needs to call a tool, interact with the environment or remain visible to an external system.
This pattern can be applied in four steps:
- Let the model optimize its internal reasoning.
- Route every tool call through a readable schema.
- Separate consequential actions from reasoning, then subject execution to a policy check or explicit confirmation.
- Log inputs, tools, outputs and confirmation status so the action can be reviewed.
Private reasoning does not mean unobservable action. Teams can optimize how a model thinks while keeping its interaction with external systems behind a clear interface.
The paper’s benchmark results and token reductions remain results reported by its authors, not evidence of production performance.
|
|
|
GOODREADS
When agents transact with agents
• This paper studies financial interactions between agents and the infrastructure needed for automated economic coordination. It expands the payments story from one agent calling an API to multiple agents that need to identify one another, grant permissions, handle failures and recover from bad transactions.
Payment agents face prompt injection
• The paper examines a specific failure mode: prompt injection can change a payment action even when the agent initially received valid authorization. It applies to any workflow that lets an agent turn user intent into a financial action.
RSI-Anything learns from selected failures
• The paper uses a world model to guide actions and select training data, while checking compatibility between a policy and a new model. Its view of self-improvement is that a system should not only update from failure, but also decide which failures are worth learning from first.
Can open models catch up with the frontier?
• The conversation with Nathan Lambert covers Chinese labs’ compute constraints, distillation, data, RL environments, reward hacking and Trillium, a lab that wants to provide environments others would usually have to buy. It connects the open-model debate to compute and training infrastructure costs rather than only comparing which model ranks first.
|
|
|
