Tech Shots
AI news flashes
עברית

AI flashes – Page 3

Sun, September 6, 2026
Models

OpenAI hits automated research intern goal

OpenAI said on September 6, 2026 that, by its own measurements, it has reached the goal announced last fall of an automated research intern by September: a system that can carry out well-defined research tasks under human direction, including work that would take a skilled researcher a few days, and that it is making strong progress toward an automated AI researcher by March 2028. The research publication reports coding agents reshaping researchers’ days—by mid-August the median researcher used more than $600/day of inference at API prices, the 90th percentile more than $7,000/day, and about 3.1 agent-workdays of effort per human workday—while stressing humans still set priorities and that full aligned recursive self-improvement is not yet known to be safe. This is an OpenAI transparency snapshot on research acceleration and RSI progress, not a ChatGPT product launch and not the GPT-6 Astra release.

Tools

OpenAI Daybreak: $1B for frontline cyber defenders

OpenAI said on 3 Sep 2026 it is launching Daybreak for Frontline Defenders, a global initiative with a $1 billion commitment in subsidized Daybreak access, training, technical support, and partnerships for resource-constrained cyber defenders—aimed to be consumed over the next six months, starting in the U.S. with expansion to partner countries in the coming weeks. The package includes Daybreak for America (priority for water/wastewater, electric grids, state and local governments, community banks, nonprofits, and open-source maintainers), a public-sector and water pilot with MS-ISAC, and more than 35 partner products in the Daybreak Defense Network; Daybreak Blue uses mainline models for common defense, Daybreak Red adds specialized cyber models for vetted orgs (OpenAI cites thousands of defenders across 2,000 approved workspaces). This is the Daybreak access program, not the GPT-6 Astra model launch and not Path to Astra.

Models

Microsoft AI ships MAI-Transcribe-2 speech recognition

Microsoft AI said on 3 Sep 2026 that MAI-Transcribe-2 is available to demo in Microsoft Foundry, the MAI Playground, and Open Router. The company says the model adds speaker diarization, word-level timestamps, keyword biasing, verbatim/clean styles, and code-switching, and claims first place on FLEURS across 60 languages (avg WER 5.2%) plus the Artificial Analysis accuracy-latency Pareto frontier — those rankings are Microsoft’s citations as of the post. Launch pricing is $0.10 per hour of audio through year-end; this is Microsoft’s STT stack, not Muse Voice Transcribe and not ElevenLabs Scribe.

Models

Microsoft opens MAI-Image-2.6 to Foundry developers

Microsoft AI said on 4 Sep 2026 that MAI-Image-2.6 — plus a faster sibling, MAI-Image-2.6-Flash — is available to developers in Microsoft Foundry and the MAI Playground. Both models add multi-image reference editing, web grounding, and dynamic aspect ratios (up to about 1.5K); Microsoft claims Flash generates images about 2.8× faster than GPT-Image-2-Medium with higher efficiency. Arena and Artificial Analysis ranks on the post are Microsoft’s own citations as of that date — this is Microsoft’s image stack, not Midjourney or OpenAI DALL·E branding.

Models

Lyria 3.5 lands in Gemini app and API

Google said on 4 Sep 2026 that Lyria 3.5, its best-sounding music generation model, is available in the Gemini app and the Gemini API, with more expressive vocals and richer arrangements. In the Gemini app users can pick or describe a genre, choose vocal or instrumental, use new templates, and choose short or longer tracks; it is available globally on web and mobile. Artists and AI creatives also get it in Google Flow Music, and developers via Google AI Studio and Google Vids — this is Lyria music generation, not Gemini video/Omni and not the Gemini crypto exchange.

Tools

GPT-6 Astra now available in GitHub Copilot

GitHub said on 4 Sep 2026 that OpenAI’s GPT-6 Astra is generally available in GitHub Copilot for Pro+, Max, Business, and Enterprise users across VS Code, Visual Studio, CLI, coding agent, Copilot app, github.com, Mobile, JetBrains, Xcode, and Eclipse (gradual rollout). In GitHub’s internal testing it stood out for planning and validating as it works, with stronger long-horizon coding and fewer steps than prior OpenAI models in Copilot. It is billed at provider list pricing under usage-based billing; Business/Enterprise admins manage access via Copilot model policy (new models on by default unless global default enablement is off or this model is blocked). This is GitHub Copilot availability, not the OpenAI GPT-6 Astra launch flash, and not Microsoft 365 Copilot.

Tools

GitHub to retire older Copilot models Oct 2

GitHub said on 3 Sep 2026 it will deprecate Gemini 3.5 Flash, Gemini 3.6 Flash, Kimi K2.7 Code, and Claude Opus 4.7 across all GitHub Copilot experiences on 2 Oct 2026, pointing users to Gemini 3.8 Flash, Kimi K3, and Claude Opus 5. Business and Enterprise admins may need to enable the replacement models in Copilot model policy.

Sat, September 5, 2026
Models

OpenAI launches GPT-6 Astra to ChatGPT and API

OpenAI introduced GPT-6 Astra on 3 Sep 2026 as its most capable model for computer use, coding, cybersecurity, science, and professional work. Rollout starts with a limited set of organizations and, over the coming days, reaches ChatGPT Plus, Pro, Business, and Enterprise plus the OpenAI API, Microsoft Azure, and AWS Bedrock (model id gpt-6-astra; Standard API pricing $10 / $50 per million input / output tokens). OpenAI says Astra saturates FrontierMath Tier 4 at 98%, ARC-AGI-3 at 99.9%, and ExploitBench at 100%, meets the Critical cyber threshold under its Preparedness Framework with stronger safeguards, and plans Daybreak access for broader defensive cyber workflows in the coming weeks; Enterprise Astra is off by default at launch.

Fri, September 4, 2026
Tools

Gemini 3.8 Flash lands in GitHub Copilot

GitHub said on 3 Sep that Google’s Gemini 3.8 Flash is available in Copilot for Pro, Pro+, Max, Business, and Enterprise across VS Code, Visual Studio, CLI, cloud agent, Copilot app, JetBrains, Xcode, and Eclipse (gradual rollout). In GitHub’s early testing it did well on complex terminal coding and recovering from actionable failures. On Business/Enterprise, new models turn on by default unless an admin disabled global default enablement or this model; introductory provider pricing runs through 31 Dec 2026.

Thu, September 3, 2026
Models

Google maps a male fruit fly's full brain

On September 3, 2026, Google Research, HHMI Janelia and collaborators released the first complete wiring map of an adult male fruit fly's brain and central nervous system — more than 166,000 neurons and 11,691 cell types. AI stitched millions of 2D slices into 3D neuron shapes; the map follows an earlier complete female fruit fly connectome and is meant as a public neuroscience resource. Three companion studies out the same day apply it to vision, taste and social behavior — this is connectomics research, not a Gemini product.

Tools

NVIDIA: RTX Spark PCs in October, plus a local PAIR router

At IFA 2026 NVIDIA said RTX Spark Windows PCs ship in October, with new Lenovo Yoga Pro 9n and Yoga 9n 2-in-1 designs and an Acer compact desktop; the superchip is a 1-petaflop RTX Blackwell GPU, up to 128GB of unified memory and a 20-core Grace CPU. NVIDIA also released a free open-source beta of PAIR, a Personal AI Router that spreads inference across idle PCs on a local network (Ollama and LM Studio; Windows, macOS and Linux). New llama.cpp and vLLM optimizations yield up to 1.9× higher local throughput, and Hermes Agent, OpenClaw and Perplexity Portable Computer are getting simpler local setup on NVIDIA hardware — RTX Spark, not Meta's Muse Spark and not xAI's Grok.

Models

WeatherNext 3: hourly global forecasts at 5 km resolution

Google DeepMind and Google Research launched WeatherNext 3, a global weather model that re-initializes every hour and ingests live geostationary satellite mosaics directly; Google calls it the most accurate global weather model to date, citing independent live evaluations by Brightband. It resolves temperature and moisture at 5 km, other surface variables at 10 km and atmospheric variables at 25 km — roughly five times sharper than WeatherNext 2's 25 km, six-hourly grid — and adds 100-meter wind, cloud cover and radiation for renewable-energy planning. It begins powering Google Search, the Gemini app, Maps, the Google Maps Platform Weather API and Earth Engine today, with up to 50% more accurate precipitation forecasts a day or more out; the data is queryable in BigQuery and Earth Engine or downloadable from Cloud Storage.

Business

NVIDIA to acquire Hugging Face for $12.93 billion

NVIDIA said on September 3, 2026 that it has agreed to acquire Hugging Face for $12,930,300,000, in a post signed by CEO Jensen Huang. NVIDIA says the platform stays open: developers keep choosing their own models, frameworks, clouds and inference providers, and NVIDIA compute will not be required to build or deploy through Hugging Face. The post counts more than 18 million users, over 3 million models, 500,000 datasets and 200,000 companies on the platform, and says NVIDIA is its largest contributor of open models and data.

Hugging Face will remain an open platform for the entire AI ecosystem.

Jensen Huang, NVIDIA
Models

Meta ships Muse Spark 1.3 for agents and coding

Meta AI Research released Muse Spark 1.3 on September 2, 2026, with stronger agentic and coding performance, rolling out in Muse Code and the Meta Model API. The update improves longer-horizon workflows, multitasking in messy threads, and coding efficiency (Meta engineers report ~20% fewer tool calls and ~25% fewer tokens vs 1.2). Reasoning modes ship now; max reasoning follows after more safety testing.

Tools

Copilot enterprises can set any default model

GitHub now lets enterprise admins pick any preferred Copilot model as the default for new conversations through managed settings, as of September 2, 2026. Teams can get their own default when the model key is marked overridable and team-mappings.json is updated; everyone else inherits the enterprise default. The feature is generally available for Copilot Business and Copilot Enterprise in the Copilot app, Copilot CLI, and Visual Studio Code — this is GitHub Copilot, not Microsoft 365 Copilot.

Tools

Copilot app and CLI now honor content exclusions

GitHub made content exclusion policies generally available in the Copilot app and Copilot CLI as of September 2, 2026. Exclusions set by enterprise, organization, or repository admins keep listed files out of Copilot context across agentic workflows. The change applies to Copilot Business and Copilot Enterprise — this is GitHub Copilot, not Microsoft 365 Copilot.

Models

Google ships Gemini 3.8 Flash Cyber via Fairwind

Google DeepMind introduced Gemini 3.8 Flash Cyber, a defender-focused model for vulnerability discovery and automated patching, available only to trusted partners through the new Fairwind Program. Fairwind pairs Flash Cyber with Google’s CodeMender harness so governments and enterprise defenders can find, verify, and fix flaws inside their secure cloud environment. Unlike generally available Gemini 3.8 Flash for agents, Flash Cyber is limited-access — and this is the Gemini model, not the crypto exchange.

Tools

NVIDIA and CrowdStrike launch SafeMind agentic cyber stack

At Fal.Con 2026, NVIDIA and CrowdStrike announced SafeMind, an agentic cybersecurity system whose defensive models are built on NVIDIA Nemotron and post-trained on CrowdStrike threat data. CrowdStrike says Blue Solano (on Nemotron 3 Super) beat leading frontier models on internal evals at far lower cost, and SafeMind ships natively in Falcon with optional QuiltWorks access to the models.

Wed, September 2, 2026
Models

Google Research ships TimesFM-3 for multivariate forecasts

Google Research introduced TimesFM-3, a 330M-parameter time-series foundation model pretrained for zero-shot multivariate forecasting on more than a trillion time points. It jointly predicts related series with past and known-future covariates in one forward pass, and ranks top among pretrained foundation models on Gift-Eval, FEV-Bench, and Time. Weights are on GitHub and Hugging Face under a non-commercial license; BigQuery integration is coming — this is TimesFM, not Gemini and not the crypto exchange.

Tools

NVIDIA announces Jetson Orin Nano 2 for robotics

NVIDIA announced Jetson Orin Nano 2, an entry-level robotics computer with twice the inference of the previous Nano in the same size, and 40% less power at the same performance in 15-watt mode. It has 78 TOPS, 8GB of memory, and an 8-core Arm CPU for robots, drones, and vision systems. The module and developer kit are expected in the first half of 2027, not on shelves now.

Tools

NVIDIA puts Groq 3 LPX inference accelerator in production

NVIDIA said Groq 3 LPX, its interactive inference accelerator, is now in full production as an extension of Vera Rubin. The company cites 3,400 output tokens per second on Gemma 4 31B at 100,000-token context, four times the nearest alternative in an Artificial Analysis benchmark.

Models

DeepMind ships Gemini 3.8 Flash for agentic work

Google DeepMind lists Gemini 3.8 Flash as generally available: its most intelligent Flash workhorse yet for coding and agents, with 1M input and 64k output tokens. It is available in the Gemini app, Gemini API, Google AI Studio, Antigravity, Gemini AI Mode, and Gemini Enterprise Agent Platform. DeepMind says it beats 3.7 Flash on long-horizon software engineering and agent benches, including 54.9% on HLE-Verified — this is the Gemini model, not the crypto exchange.

Tools

GitHub Copilot in VS adds thinking effort controls

GitHub added Low, Medium, and High thinking-effort controls in Visual Studio for supported models, plus org-level custom agents in the picker. You can pin models, see context window and cost, and ask the Git agent to review uncommitted changes before a PR, including Azure DevOps repos. It is on every Copilot plan including Free.

Tools

Power Apps canvas authoring agent plugin hits GA

Microsoft made canvas apps coauthoring with agents generally available on September 1: makers and coding agents share one authoring model through an MCP-backed plugin that can create screens, wire connectors, write Power Fx, and sync to a live session. The output is still a normal canvas app under existing connectors, governance, and deployment—you pick what to delegate versus edit by hand to manage tokens.

Tools

OpenAI backs California SB 1119 youth AI safety

On 31 Aug OpenAI said it supports California Senate Bill 1119 for age-appropriate AI safeguards for teens while keeping educational access. The company lists requirements it backs: age determination, pre-release risk work, independent audits, protections against high-risk content, parental tools, crisis-resource links, and limits on targeted ads for ages 13–17 applied automatically. It ties the bill to ChatGPT for Teens defaults and asks Governor Newsom to sign.

Tools

GitHub Copilot VS Code August: agent sessions, /btw

GitHub’s Aug 31 changelog for VS Code 1.132–1.135 adds side-by-side agent chats, /btw side conversations that share context and prompt cache, a prompt timeline, portable Agent Plugins 1.0, multi-window Agent Host sessions, and experimental /rubber-duck second opinions. Chat gets full-transcript search and sticky prompts; the integrated browser can annotate page elements for the agent; on-device dictation adds multi-language and shell-aware cleanup.

Models

DeepMind pilots double-blind frontier AI evals

Google DeepMind said on 27 Aug it is piloting what it calls the first double-blind evaluation of a proprietary frontier model: external tests run inside a cryptographic Confidential Space box so evaluators cannot see Gemini weights and Google cannot see the evaluators’ prompts. Partners include the Singapore AI Safety Institute, OpenMined, AVERI, and MLCommons, starting with a Gemini Flash Lite model on confidential benchmarks. The goal is less benchmark contamination and stronger trust for sensitive safety tests — not a new Gemini consumer feature, not Omni video, and not the Gemini crypto exchange.

Tools

Claude Fable 5.1 lands in GitHub Copilot

GitHub said on 1 Sep that Claude Fable 5.1 is generally available in Copilot for Pro+, Max, Business, and Enterprise across VS Code, Visual Studio, CLI, coding agent, github.com, Mobile, JetBrains, Xcode, and Eclipse (gradual rollout). Business/Enterprise admins must enable a policy that is off by default. Unlike most other Claude models in Copilot, Fable 5.1 retains prompts and outputs for Anthropic safety classifiers by default (not for training); eligible enterprises can get time-limited zero data retention until EFS.

Tools

Photoshop ships AI Assisted Editor beta and Instruct Edit

Adobe said on 27 Aug that Photoshop’s AI Assisted Editor is in beta — a dedicated prompt-based workspace with Prompt to Edit, Generative Fill/Expand/Upscale, Remove, and AI Markup. Instruct Edit with Masks (Firefly Image 5) lets you describe edits while protecting unmasked faces, logos, or brand assets; Prompt to Edit also lands in the Pro Editor’s contextual task bar as generative layers.

Tools

Antigravity Teamwork pairs agents with Gemini 3.7 Flash

Google said on 31 Aug that Antigravity Teamwork — a framework for autonomous multi-agent teams that collaborate over hours or days — paired with Gemini 3.7 Flash accelerated long-horizon math and engineering work. The post cites seven open TCS problems solved (including Knuth’s Cycles Conjecture verified in Lean), a cycle-accurate RISC-V simulator that boots xv6, and upstream Eigen and ParlayHash wins.

Tools

Anthropic opens Model Hardware Standard research preview

Anthropic opened a research preview of the Model Hardware Standard (MHS), a shared, model-agnostic spec so AI agents can operate programmable lab and manufacturing devices through protocols such as MCP. Early partners span biotech, microscopy, and quantum (including Genentech, HHMI Janelia, and QuEra); Anthropic plans to open-source MHS after the preview and is building physical-safety evaluations. Waitlist only for now — research preview, not a general Claude feature for everyone.

Tools

Google Pics brings AI image edit into Docs and Slides

Google is rolling out Google Pics, an AI image creation and editing tool built on its Nano Banana model, to Google AI Pro and Ultra subscribers and most Workspace business customers. Features include object isolation, in-image text edit and translate, multi-option generations, and collaboration; Docs and Slides integrate today, Drive follows in coming weeks. Try pics.new — this is Workspace Pics, not Gemini video generation.

Tools

ChatGPT Ads hits $1B ARR; Ads Manager expands

OpenAI says ChatGPT Ads reached a $1 billion annualized revenue run rate in under 200 days after launch, with tens of thousands of advertisers. Starting 31 Aug, advertisers can buy ChatGPT ads directly via Ads Manager across India, Europe, the Middle East, and North Africa; Israel is not on the official self-service country list. Ads sit alongside subscriptions, enterprise, and API revenue, and fund an ad-supported free tier for more than 1 billion weekly active users.

Tools

GitHub Copilot code review can approve pull requests

Copilot code review now includes an approval assessment on every overview comment, and admins can optionally let Copilot submit approvals that count toward required reviews (off by default). Controls sit at enterprise, org, and repo level, including path filters; new commits dismiss the approval like a human review. Public preview for Copilot Pro through Enterprise — this is GitHub Copilot, not Microsoft 365 Copilot.

Tools

ChatGPT for Healthcare connects Epic EHR and public data

OpenAI added an Epic EHR integration to ChatGPT for Healthcare so clinicians can pull authorized chart context into ChatGPT or run ChatGPT inside the EHR layout in supported deployments. A new Healthcare Public Data plugin adds structured connectors to nine official sources, including PubMed, DailyMed, CMS Coverage, ClinicalTrials.gov, and RxNorm. ChatGPT for Healthcare customers ask a workspace admin to enable both; Enterprise customers confirm eligibility with OpenAI, with enterprise controls for governed clinical workflows.

Models

Meta launches Muse Voice Transcribe for real-time ASR

Meta Superintelligence Labs shipped Muse Voice Transcribe, its first real-time audio perception model: streaming speech-to-text, diarization for 20+ speakers, endpointing, and adaptive delay. It is trained on 70+ languages (25 extensively verified at launch) with in-sentence code-switching, and Meta says it ranks first on Artificial Analysis streaming STT as of 1 Sep. Live today via the Meta Model API, Meta AI for Mac (hold Fn to dictate), and Muse Code.

Models

Gemini adds agentic video understanding in the API

Google launched agentic video understanding on 1 Sep across Gemini 3.7 Flash, 3.6 Flash, and 3.5 Flash-Lite: the model searches and inspects segments instead of ingesting a fixed frame rate. The company says that cuts token use by up to 88% and cost by up to 66%, with accuracy up about 7%; it's in the Gemini API in Google AI Studio and Gemini Enterprise today, at standard token prices with no extra feature fee.

Models

Anthropic launches Claude Fable 5.1 and Mythos 5.1

Anthropic launched Claude Fable 5.1 (generally available) and Mythos 5.1 (trusted-access only for cybersecurity and life sciences)—same underlying model, different safeguard tiers—aimed at coding, knowledge work, and agentic research. The company says typical token-billed workloads cost about 25% less than Fable 5 via cheaper cache reads, and up to about 45% less on highly agentic work. Enterprise Frontier Safeguards (customer-controlled storage plus misuse detection) rolls out later this fall; eligible customers keep zero data retention until then.

Models

OpenAI: Astra crosses Critical cyber threshold

OpenAI said on 1 Sep that its forthcoming model Astra is the first to cross the Critical cybersecurity capability threshold in its Preparedness Framework: with tools and access it can find unknown flaws and develop exploits across many well-protected systems without step-by-step human guidance. Release is still planned "soon," but the most advanced cyber capabilities start with a limited tester group, then Daybreak Blue for defensive use. OpenAI says it delayed parts of Astra's development and release while it strengthened and tested protections against cyber misuse and unauthorized model actions, and now believes those safeguards sufficiently minimize severe-harm risk for release under the framework.

Tools

Adobe Firefly audio tools now generally available

Adobe said on 20 Aug that Generate Music, Generate Speech, and Generate Sound Effects are generally available in Firefly. The company says Generate Music makes universally licensed original tracks it calls commercially safe; Generate Speech runs on the Firefly Speech Model, with an option to use ElevenLabs — Adobe did not say ElevenLabs licensing is covered by Firefly indemnity.