News

Accessing OpenAI models on Amazon Bedrock from Australia with global cross-Region inference

Australian teams can now call OpenAI's GPT-5.6 Sol, Terra, and Luna models on Amazon Bedrock from the Asia Pacific (Sydney) and Asia Pacific (Melbourne) Regions. That access runs through global cross-Region inference, so the workloads do not need to move to another geography. All three models take text and image inputs, generate text, and carry context windows that reach 1 million tokens.

A full production path, from invocation to CloudWatch

OpenAI fits the three models into different roles. Sol is aimed at demanding reasoning, coding, and agentic workloads, while Terra balances performance and cost for everyday production use. Luna targets fast, affordable inference for high-volume and latency-sensitive applications. The same post walks through calling the models, using prompt caching, and setting up Codex with OpenID Connect authentication. It also shows how to monitor usage with Amazon CloudWatch. The availability spans both named Australian AWS Regions rather than one. (AWS Machine Learning)

thinkidiot take: 1 million tokens of context is the number that changes what you can queue up in one call, and a Sydney-based team finally gets it without the round trip. If I am running agentic jobs on Sol, I stop splitting work across calls just to dodge a window limit, and prompt caching does the rest. That pushes a capability that historically required US-based infrastructure into the same toolchain as local governance, billing, and CloudWatch monitoring. It also undercuts the argument that Australian data residency and frontier-model quality are one or the other. The practical read is that the model layer and the compliance layer finally share a vendor in Australia.

Proactive cyber defense for governments and enterprises

Google is launching the Fairwind Program, a limited access program that gives governments and trusted partners use of its advanced cyber defense capabilities. The program pairs the Gemini 3.8 Flash Cyber model with the CodeMender harness. Together they cut the time to fix a vulnerability from weeks of manual work to minutes that produce verified, deployment-ready patches.

Defense as a named program, not a product page

Fairwind is limited to governments and trusted partners, and more than 650 participating partners globally take part. The group includes governments, critical infrastructure operators, and core technology platforms. The Gemini 3.8 Flash Cyber model is described as Google's most advanced cyber model. CodeMender delivers specialized reasoning to write and validate code fixes at a fraction of the operating cost of traditional frontier models. The workflow the program targets replaces manual vulnerability repair with generated fixes that are ready for deployment. (Google DeepMind)

thinkidiot take: Minutes instead of weeks is the claim that matters, and it lands where the cost was always felt. The patch pipeline is the slow part of a security response, and that is now where a named model and a named harness sit. Pairing that with a fraction of the operating cost of traditional frontier models means the expensive step of verification no longer demands a frontier model to get through. That reframes how a critical infrastructure operator staffs a response, since the bottleneck is no longer a single specialist's week. The limited access model also signals that Google sees defense as a relationship, not an API, and I think that is the right call for the buyers involved.

Real-Time Intelligence with IBM Time Series Models on Confluent

IBM and Confluent are bringing time series foundation models, or TSFMs, to Confluent Cloud, where the data already moves. A TSFM trained once across vast, varied signals generalizes to a series it has never seen. Access opens on Confluent Cloud on AWS first, with Confluent Platform following for on-premises and hybrid environments.

Specialist work moving onto the domain expert's desk

The models sit where the data already streams rather than requiring a separate inference cluster. IBM and Confluent describe every point of accuracy as worth millions, with productivity gains of 5 to 10 times. Work that waited for specialists now sits with domain experts. The models carry more than 44 million downloads behind them. Confluent Cloud hosting and Flink integration form the delivery layer. The on-premises and hybrid path follows after the initial cloud rollout. (Hugging Face)

thinkidiot take: The 5 to 10 times productivity gain is the number that changes the staffing plan, and it lands with the people who already understand the signal. That puts the model next to the domain expert instead of behind a specialist queue, which is where the delay has always been. Training once and generalizing to an unseen series also drops the model build from a project to a default option. It changes what a Confluent deployment is for, from moving data to reasoning about it. The cloud-first, hybrid-following rollout tells you the buyers are being met where the data already lives, and that is the right order of operations.

An Organizational Second Brain: Building an AI That Learns From Experts

Meta has built an AI agent that acts as a secondary expert for a given domain. It integrates a structured, auditable knowledge architecture with a self-improvement loop. The agent is making deep specialist knowledge available to anyone in an organization to access, share, and build upon.

Expertise that survives the specialist leaving

Most valuable specialist knowledge lives in people's heads and rarely gets captured durably. The same types of questions keep arising across hundreds of product reviews in compliance domains. Expert assessments take days of manual research, and inconsistency between assessments creates real organizational risk. Experts spend more time answering routine questions than on genuinely novel and ambiguous work where their judgment matters most. The system is already saving domain subject matter experts at Meta substantial time. That time is redirected toward the work where their knowledge matters most. (Meta Engineering)

thinkidiot take: The days of manual research for a single expert assessment is the cost that this system removes, and it removes that cost at the point where inconsistency was the real risk. The knowledge architecture plus self-improvement loop is durable because the value comes from the structure holding while specialists rotate. If the same compliance question has shown up across hundreds of reviews, the answer should have stopped being a specialist's week. Meta treats the expert's head as the source of truth and the model as the copy, which is the right direction , this is the first time an organization-scale secondary expert has the auditable backbone to actually be relied on.

The Builders Stage brings practical strategies for scaling startups to TechCrunch Disrupt 2026

The Builders Stage is returning to TechCrunch Disrupt 2026. It is one of six industry-focused stages at the event. More than 10,000 founders, investors, startup operators, and technology leaders are set to attend at Moscone Center in San Francisco from October 13 to 15.

Six stages, one built for the growth decisions

Builders Stage is dedicated to founders navigating the challenges of growth. The coverage stretches from raising capital and hiring to go-to-market strategy and AI. Founders, startup operators, and investors meet for practical conversations on what it takes to build and scale. Every session delivers strategies that can be put to work immediately. Live Q&A gives a direct line to the speakers. Registration is open today, with rates rising and savings of up to 330 dollars on the table. (TechCrunch)

thinkidiot take: Ten thousand people in one room in San Francisco is the scale that changes what a conference stage is for. The value of a Builders Stage session is not the talk, it is the live Q&A that turns a general strategy into a specific call for your company. That is where a founder gets an answer about their capital raise or hire, not a slide about it. The 330 dollar savings is small next to a week of misdirected effort, and the six industry-focused stages at Disrupt 2026 let a founder pick the room for their actual problem.

Trending AI Papers

Ranking source: Hugging Face Papers for 2026-09-03.

Repo-To-Skill: Distilling GitHub Repositories Into AI4AI Skills

AI agents can know a research method without knowing the practical steps that make it work. This paper turns useful material from software repositories into compact instructions that an agent can consult during a task. Its goal is to help research agents reuse proven know-how instead of rebuilding it each time.

• Problem: Research agents have tools for planning, execution, memory, and checking results, but they still lack much of the practical knowledge needed in specific fields. That knowledge exists in papers and repositories, yet it is written for people and is often too large to give an agent during a task.
• New idea: The authors built DisCo, a research agent that creates and applies compact instruction sets called skills. A skill captures the practical steps needed to perform a particular capability. DisCo can extract broadly useful skills from popular repositories or create skills for the needs of a specific task. The broader process produced the AREX-Skill Library, a collection organized by research area and capability.
• Simple example: It is like giving a cook a set of tested recipe cards instead of asking them to search an entire library whenever they prepare a meal.
• Evidence: The library contains more than 5,000 verified skills from 1,000 widely used ML repositories, covering 20 areas and 178 capability families. With the GPT-5.5 backbone, research harness, and execution budget kept fixed, adding skills raised scores by 134.3% on MLE-bench, 34.4% on PaperBench, 9.2% on FrontierCS, and 14.0% on PassNet.
• Limitation: The abstract does not test whether the gains carry over to other model backbones, benchmarks, or fields outside ML research.
• Why it matters: Compact practical knowledge can make research agents more capable without changing their underlying model or execution budget.
• Paper: Repo-To-Skill: Distilling GitHub Repositories Into AI4AI

SolarWM: Open Data and Scalable Training for Long-Horizon Video World Models

Video world models try to generate a changing visual environment that a user can influence. Building them is hard because video collections and model designs represent scenes and motion in different ways. SolarWM provides an open system for preparing varied data, training several model families, and running them for long periods.

• Problem: Video datasets differ in timing, camera information, image quality, movement, and caption style, while generators use different internal designs. Simply combining the data or writing separate code for each model creates uneven training signals and makes experiments hard to reproduce or compare.
• New idea: SolarWM uses a multi-source data engine, a system that converts material from different datasets into one common format. That format aligns video frames with camera measurements, captions, quality information, selection records, and source history. A backbone-native adaptation framework then connects this shared data and training setup to each video model while retaining the model's original representation and objective. Its training recipe moves through bidirectional adaptation, initialization from guided next-step generation, and distillation that matches the teacher model's output distribution.
• Simple example: It is like translating footage from many cameras into one editing format, then giving each projector the adapter it needs without rebuilding the projector.
• Evidence: The engine converts 1.43 million clips from 10 datasets into a shared format. The authors built four models ranging from 5B to 33B parameters using Wan2.2, LTX-2.5, and MiniMax-H3. After training on sequences only 5 seconds long, the causal models support real-time interaction over generated runs lasting from minutes to hours.
• Limitation: The abstract does not report comparative quality results or show how reliably the models preserve coherent scenes and actions during rollouts lasting minutes or hours.
• Why it matters: An open common foundation could make interactive video-model experiments easier to reproduce, compare, and extend.
• Paper: SolarWM: Open Data and Scalable Training for Long-Horizon

EarlyEval: Cheaper Agent Evaluation via Early Outcome Prediction

Testing an AI agent can be expensive because each benchmark task may require a long sequence of model calls. EarlyEval tries to stop runs once their eventual result is already clear from what the agent has done so far. This reduces evaluation work inside each task rather than removing tasks from the benchmark.

• Problem: A single benchmark pass with a frontier model can cost hundreds to thousands of dollars, and developers repeat that expense across many iterations. Earlier methods shrink the set of test tasks but do not reduce the cost of running each task that remains.
• New idea: Early outcome prediction estimates whether an agent will succeed or fail before its run finishes. EarlyEval implements this with two LightGBM classifiers, which are lightweight prediction models built from decision trees. One predicts success and the other predicts failure using behavior, text, and features from a reference solution. The framework stops the run when either prediction exceeds a confidence threshold chosen to control reliability.
• Simple example: It is like ending a driving test early when the examiner already has enough evidence to know whether the driver will pass or fail.
• Evidence: Across SWE-bench Verified, TerminalBench, and Toolathlon, EarlyEval removed 13% to 26% of agent steps, up to 44.1% of input tokens, and up to 29.4% of output tokens. Its predictions were 89% to 97% accurate, while average resolve rates changed by only one to two percentage points per agent.
• Limitation: Stopping early still changes average resolve rates by one to two percentage points, and the abstract does not establish performance beyond the three tested benchmarks.
• Why it matters: Cheaper evaluations could let developers test agent changes more often without greatly altering the measured results.
• Paper: EarlyEval: Cheaper Agent Evaluation via Early Outcome

Trending AI Repositories

Ranking source: GitHub Trending.

google-research/timesfm

Google Research has released a new TimesFM checkpoint for forecasting time-series data. The project is also backed by an ICML 2024 paper, making it worth a fresh look.

• What it is: It is a Python project with model checkpoints hosted on Hugging Face. The listed checkpoints include the new TimesFM 3.0 PyTorch release and earlier versions up to 2.5.
• What it does: TimesFM (Time Series Foundation Model) is a pretrained time-series foundation model developed by Google Research for time-series forecasting.
• Who it helps: It helps researchers and developers working on time-series forecasting. They can use the published checkpoints and consult the accompanying paper.
• Limitation: Using the new checkpoint requires PyTorch.
• Repository: google-research/timesfm

DietrichGebert/ponytail

Ponytail is a JavaScript project built around a deliberately minimalist coding philosophy. It is worth a look for teams that want agents to solve problems with less code.

• What it is: It is a JavaScript repository for shaping how an AI agent approaches coding tasks. Its stated style favors saying little and writing a single working line.
• What it does: Makes your AI agent think like the laziest senior dev in the room. The best code is the code you never wrote.
• Who it helps: It helps developers who use AI agents to write code. They can push those agents toward smaller, simpler solutions.
• Limitation: Its minimalist approach may not suit tasks that require substantial code or explanation.
• Repository: DietrichGebert/ponytail

ChromeDevTools/chrome-devtools-mcp

This TypeScript project connects AI coding assistants to a live Chrome browser. It matters because agents can inspect real browser behavior instead of reasoning from code alone.

• What it is: It is an MCP server that sits between a coding assistant and Chrome DevTools. It exposes browser control, inspection, automation and debugging capabilities.
• What it does: Chrome DevTools for coding agents
• Who it helps: It helps developers using agents such as Antigravity, Claude, Cursor or Copilot. Their assistants can automate Chrome and investigate browser problems in depth.
• Limitation: It requires a live Chrome browser and an AI coding assistant that can use MCP servers.
• Repository: ChromeDevTools/chrome-devtools-mcp