Curating all things AI.
Joined March 2026
- Tweets626
- Following68
- Followers86
- Likes789
Lukasz Olejnik says he singled out half a million @Google searches across four countries in August 2026 under the EU's anonymisation scheme.
The EU requires sharing with rivals, including AI chatbot makers.
→ Browser display language was not covered by the anonymity checks
→ In Germany that month, he linked 200,000+ sets beyond the three-search cap
Google paid for the audit. Olejnik says he had full independence.
1
4
🚨BREAKING: NVIDIA says @JensenHuang has resigned as president, CEO and board member, effective immediately.
→ CFO Colette Kress becomes interim CEO and keeps her finance role until a permanent successor is named
→ The board is starting a permanent CEO search and will consider internal and external candidates
→ Technology executive Harvey C. Jones becomes the board's first-ever chair
1
1
51
Microsoft’s FORGE Lab says AI-assisted Linux bug hunting yielded 59 upstream reports, but preparing them took over four person-months.
Two weeks of scanning selected kernel subsystems produced 26,502 raw findings. Reproduction agents narrowed the pool to 621 cases with proof-of-concept evidence for expert review.
→ Eight of the 33 assigned CVEs had working exploits that gained local administrator access in lab tests
→ Linux maintainers have merged fixes for 38 reports
→ The reported bugs averaged 8.1 years old
→ Merged fixes took a median 41 days from first public patch to a mainline release
Some exploits required non-default settings or disabled exploit protections.
microsoft.com/en-us/security…
10
AWS just launched a Web Search API on Amazon Bedrock AgentCore, letting code search the web with one API call.
Previously, developers had to set up a gateway and search target.
→ Call it through @awscloud SDKs or the CLI with existing AWS credentials
→ Results include ranked text extracts from Amazon's own web index
→ Filter by domain and publication date
Page titles, URLs and publication times are not always available.
aws.amazon.com/blogs/machine…
24
Matthew Goldberg says agents wrote 100% of the code in the newly released AWS Emulator for building and testing cloud applications locally.
→ Downloadable binary, with CC0 source and tests on GitHub
→ Tens of thousands of tests compare its behavior with real AWS
→ Thousands of agents write tests, find issues and fix code
No human reviews the code. Automated gates check security and test for regressions.
aboutamazon.com/news/aws/how…
20
Microsoft says GPT-6 Astra costs about 5× as much in completion tokens as GPT-6.1 Sol in its document tests.
Azure Content Understanding now supports GPT-5.6, GPT-6 and GPT-6.1.
→ Start with Sol for standard document extraction
→ GA schema-only extraction accuracy is 95.95% for Astra versus 95.74% for Sol
Results vary by task. Test on your own documents.
devblogs.microsoft.com/found…
1
34
Google says its new API scores up to 1,000 questions about one input in a single pass.
Gemini Decisions API is available in @Google AI Studio. Gemini Decision 1 returns probabilities, not generated text.
→ Supports text, images, audio, video and PDFs
→ Google reports 150–250 ms end-to-end latency
→ $0.10 per 1M input tokens, with no output tokens billed
Limited to predefined choices and inputs up to 256K tokens.
blog.google/innovation-and-a…
10
Google Labs introduces Tabletop, shared spaces where groups can build and use custom planning tools together.
→ Add links, files and ideas, then prompt @tabletop or pick a starter template
→ Generate interactive trip itineraries, dinner polls, expense splitters or tournament brackets
→ Share an invite link so everyone can use the space at once
Rolling out in the U.S. for ages 18+ via waitlist.
blog.google/innovation-and-a…
15
AWS just added metering so speech AI providers can bill for actual usage during live streams on Amazon SageMaker AI.
Deepgram, a speech AI company, is the launch partner for transcription and speech generation billing on @awscloud Marketplace.
→ A separate WebSocket connection carries usage records
→ Metering is capped at one message per second per connection
SageMaker AI endpoint infrastructure charges still apply.
aws.amazon.com/blogs/machine…
23
Adam Gayoso and coauthors say their new model predicts how DNA changes affect gene activity across hundreds of human cell types.
AlphaGenome 2 moves beyond its predecessor’s broad tissue averages to specific cells and states.
→ Can predict effects in disease-relevant cells that broad-tissue models miss
→ Can make predictions for cell and disease states unseen during training
deepmind.google/blog/alphage…
22
As @nvidia’s alleged honeypot scandal unfolds, @AlibabaGroup just disclosed an earlier confidential Hong Kong IPO application for T-Head, its data-center chip business.
→ The proposed offering would give T-Head independent access to equity and debt markets, while it is expected to remain an @AlibabaGroup subsidiary
→ Alibaba says Zhenwu AI chips serve over 650 customers. Its SAIL software includes compilers, libraries and multi-chip communication tools
→ The Hong Kong exchange has allowed the spin-off to proceed, but final listing approval is still required
→ Offer size, structure and dilution remain undecided, and Alibaba shareholders have no guaranteed share entitlement
The listing also requires completion of the Chinese regulatory filing and final company decisions.
1
14
NVIDIA says Alpamayo 2 Edge drives about 60× farther between at-fault incidents than Qwen-Drive-1.0 in AlpaSim.
The 2B-parameter model turns camera views and navigation inputs into driving trajectories onboard DRIVE AGX Thor. It is available on Hugging Face.
→ @nvidia reports 160–190 ms from camera pixels to a planned trajectory on Thor
→ The model generates reasoning only when needed to save compute
→ Closed-loop reinforcement learning in AlpaGym improves driving performance by 48.9% over the best checkpoint without that training, NVIDIA reports
→ Developer Hub recipes cover quantization and inference acceleration with TensorRT and DFlash2
Latency depends on how much reasoning the model generates.
developer.nvidia.com/blog/ru…
46
OpenAI says its new business profiles and agents connect businesses with ChatGPT's 1.2 billion weekly users.
Businesses can claim and verify their profiles through Business Hub, now in beta.
→ Verified businesses can add product feeds with pricing and availability
→ Aggregated insights show mentions and website clicks without sharing users' conversations or Business Agent chats
→ Customers can start a separate, labeled Business Agent conversation from a business's profile
→ Businesses can sponsor their agents as ChatGPT Ads destinations, with placements clearly labeled and separate from ChatGPT responses
Business Agents are limited to eligible ecommerce and online service businesses in the United States.
openai.com/index/introducing…
15
MiniMax released H3-Image, an image generation model it says can produce a 4K (4096 × 4096) image in under 10 seconds.
The weights are on Hugging Face for customization, with hosted access through the @MiniMax__AI API Platform, MiniMax Code and Hailuo AI. It supports text-to-image generation, reference-based generation and image editing.
→ Reference-based generation supports up to 16 images
→ MiniMax reports 1.5–2× the speed of mainstream alternatives in Standard mode and 3–4× in Highspeed mode
→ The company puts pricing at 50–70% of comparable mainstream offerings
→ Launch-week requests at 4K resolution and above get 50% off the standard price
The speed and price comparisons are vendor-reported, and MiniMax does not name the alternatives.
minimax.io/blog/minimax-h3-i…
47
NVIDIA says its new Snapshot component starts and restores GPU workloads in seconds.
Starting an LLM workload from scratch can take minutes.
→ It reuses initialized GPU workloads across inference replicas to speed startup and recovery
→ The component is open source and part of the @nvidia DSX AI Factory Platform
developer.nvidia.com/blog/st…
16
Anthropic launched its IPO roadshow with 825 million shares on offer at an expected $70–78 each.
The company has applied to list its Class A common stock on the Nasdaq Global Select Market under the symbol ANT.
→ @AnthropicAI is offering 801,351,354 shares
→ Existing stockholders are offering 23,648,646 shares
→ The company will receive no proceeds from the stockholders' sales
→ Anthropic expects to grant underwriters a 30-day option to buy up to 123,750,000 additional shares
The SEC registration statement is not yet effective. Shares cannot be sold, nor offers to buy accepted, until it becomes effective.
anthropic.com/news/roadshow-…
34
Google Cloud launched Gemini for healthcare to connect patient records and handle administrative tasks.
Google says clinicians lose nearly 28 hours a week to paperwork and administrative tasks.
→ It links electronic health records, billing, clinical guidelines and Medicare data
→ Tasks include medical records summaries and prior authorization review
Preview only, with defined guardrails and human oversight.
blog.google/innovation-and-a…
26