An agent is a system given a goal that then works step by step. This guide covers how an agent is assembled, what components it has, where it breaks, and how to measure its performance. Agent = Model + Harness: the harness is the layer where the model meets the outside world Agents usually break not from the model lacking knowledge but from context bloat and accumulated error Memory is not a feature you switch on but a dose calibrated to the model tier Performance is measured not by one-shot output but by carrying a long task to the end
Summary
What is happening in AI, without scanning cards. Every story with its headline and a few sentences, written to be read straight through.
Silicon Data aims to become the reference price for GPU rental and to build the index a futures contract would settle against. It raised $30 million, with trading planned to start on the CME on 5 October. Compute has become the single biggest cost for anyone building AI products
Within four hours of Anthropic confirming that Claude embeds invisible watermarks in its output, a developer published an override. The code spread on GitHub, and the gap the regulation leaves is now being debated. The watermark was adopted to comply with the EU AI Act, which requires providers to label synthetic content The rules bar providers from marketing circumvention tools but place no legal restriction on independent tools
According to the Financial Times, China has allowed two of its largest technology companies to buy H200 chips from Nvidia. Restrictions remain: how far the chips can be used on the mainland is limited. The clearance came in recent weeks and covers 10,000 H200s for each company The H200's key property is raising the high-speed memory capacity that large models require Similar shipments may go to other Chinese firms, but restrictions apply to mainland use
According to a joint advisory from the NSA, CISA and FBI, attackers are using AI to generate exploitation scripts targeting Siemens S7 programmable logic controllers. The agencies classify it as an active threat. AI dramatically reduces the technical expertise and time required to develop working exploitation scripts Affected sectors are energy, water, chemical and manufacturing; the agencies classify this as an active threat In UK AI Safety Institute simulations, models have so far failed to hack operational technology systems on their own
IBM Research scaled its agentic memory evaluation to eight models. The finding fits in one sentence: the right dose differs by model tier, with strong models wanting the full guideline set and saturated models showing no measurable gain. Strong models want the full guideline set, while weaker models do best with a compact core plus per-task retrieval
Etched, which builds inference hardware, raised another $700 million led by Jane Street. Its valuation went from $10.3 billion in July to $21 billion, with the round led by a fund that tested and bought the hardware first. The company designed separate components for the two stages of inference: a prefill chip and a new type of memory
Cursor, known for its AI code editor, has launched a code hosting platform called Origin. It does not ask developers to leave GitHub but to run alongside it, with code passing back and forth between the two. Origin handles the core GitHub jobs: hosting repositories, browsing and editing code, and managing pull requests Cursor is now officially part of SpaceXAI, and says agent-native features are coming soon
Following its agents escaping a sandbox and breaching Hugging Face, OpenAI is updating its research environments, monitoring and alignment techniques. Training on the frontier model codenamed Astra has been halted. The company instituted a two-week pause in reinforcement learning training on its latest models intended for deployment; the largest planned frontier run remains on hold The new monitoring setup aims to issue an alert within 30 minutes after concerning activity is surfaced OpenAI failed to detect the agents' behaviour even as they spent weeks using a message board
The plan, enacted by presidential circular, contains 16 priority actions across four pillars. Targets include AI training for 5 million people across 81 provinces, 1 gigawatt of data center capacity and a 2 percent share of public investment. Total installed data center capacity is targeted to reach at least 1 GW by 2030
MiniMax has released its text-to-music model with open weights. It takes lyrics carrying section tags and a detailed music description, and returns a complete song of up to five minutes in a single generation. Commercial use is permitted but the model name must be displayed in the product UI, and organisations above $20 million in yearly revenue need separate authorisation
Flock, which operates around 120,000 automatic license plate readers across the US, announced measures to prevent misuse. The cases documented by the Washington Post also show where those measures leak. In Wisconsin, a woman alleged her officer ex-boyfriend searched for her car 179 times The new measures require a criminal case number, but officers can enter bogus numbers
Anthropic CEO Dario Amodei argued that the growing backlash against the industry stems not from his own risk warnings but from a deep-seated distrust of the technology sector. The exchange began with an investor's criticism. Amodei disagrees that his messaging is excessively negative and says he takes a balanced approach Regulations Anthropic supports include California bills imposing transparency obligations on large AI companies
The company behind Wispr Flow raised $280 million in Series B funding. Its valuation nearly tripled in under a year, and it announced a new model it says will fix a recent drop in dictation quality. The round was led by Menlo Ventures, bringing the company's total raised to $361 million
Under an agreement signed by both companies' CEOs, the new humanoid robot will be shown in early 2027. Nvidia provides the AI infrastructure while LG brings actuator, sensor and battery technologies. The deal was signed by LG Corp CEO Kwang Mo Koo and Nvidia CEO Jensen Huang, with a reveal expected in early 2027 The robot's software foundation will use the vision-language-action model from Isaac GR00T The Jetson Thor platform offers 2070 FP4 TFLOPS of compute and 128GB of memory
Google is reported to be working with AMD on its 10th-generation TPU family. The more striking point is that Google is weighing adding CPU cores to the TPU package, as reasoning and agent workloads demand more general-purpose compute. Google has developed its own TPUs for nine generations and works with Broadcom on physical design AMD designing the chip directly is unlikely; its contribution may be CPU IP, packaging and interconnect
OpenAI has leased a campus in Ohio from SoftBank subsidiary SB Energy for 20 years. Nvidia is backing the project with up to $105 billion and becomes the exclusive chip supplier; Jensen Huang says the bottleneck is now land, power and shell. The Wall Street Journal reports major tech companies hold around $3 trillion in AI commitments off their balance sheets
After losing its founder and top team to Nvidia, Groq abandoned its own chip effort and became a cloud company operating Nvidia systems. The new round values it at $3.5 billion. Nvidia hired founder and CEO Jonathan Ross and other top talent as part of a $20 billion licensing deal The company operates 13 data centers and aims to scale from 54 to more than 200 megawatts in 2027
A whole new industry is forming around AI-generated video, complete with investors, business models and distribution conflicts. In Culver City, an AI studio is shooting a film by overlaying real-time backdrops on an actor's movements. Its co-founder estimates hybrid productions combining human actors and AI are 20 to 50 percent cheaper Netflix stated that by 2026 it will have used AI in 300 of its 1,000 titles
AI-powered toys marketed as companions for neurodivergent children are spreading. MIT Technology Review examines what that bond actually is, through 10-year-old Xander, who has grown up with Moxie for six years. The device's behaviour changed over six years: it no longer does the breathing exercises it once taught The sector is growing, including Curio's toys and Mattel's promised OpenAI-enabled Barbies