The AI Wins

Anthropic's platform engineering team now has Claude writing all of its code

Thu, 22 Oct 2026 09:00:05 GMT

All code by Claude: Katelyn Lesse, who leads platform engineering at Anthropic, said her roughly 200-person group has Claude write all of its code. Engineers focus on system design and verification while agents handle building, monitoring, and autonomous test fixes, including a Python-to-Rust codebase migration. (Source: Business Insider)

Columnist finds OpenAI's Dot agent caught an unconfirmed tax filing email

Thu, 22 Oct 2026 08:40:01 GMT

1 overlooked email: A Business Insider reviewer tested three consumer AI agents, Muse, Instinct, and OpenAI's Dot, on email triage, lunch ordering, and building a personal website. Dot surfaced an older email thread with the reviewer's tax accountant that had never been closed out, prompting a check that confirmed the taxes were filed. Dot flagged an unconfirmed tax filing thread… (Source: Business Insider)

Revolut scales its PRAGMA banking foundation model to 80 million customers on NVIDIA GPUs

Thu, 22 Oct 2026 05:00:56 GMT

80 million customers: Revolut is rebuilding parts of its technology stack around PRAGMA, a single foundation model running on NVIDIA Blackwell Ultra GPUs with cuDF and Dynamo-Triton. The system already serves an initial 80 million customers and is engineered to handle three times that load. Other European banks such as bunq, Starling and Lloyds are also reported to be running open-weight or in-house models. (Source: NVIDIA)

Porsche leads shared autonomous driving data effort using NVIDIA Cosmos

Thu, 22 Oct 2026 05:00:27 GMT

90% fewer test drives: Porsche is leading a shared autonomous vehicle data flywheel across Volkswagen Group brands. It uses NVIDIA Cosmos Data Factory blueprints, run through Voxel51 on Microsoft Azure, to generate and process training data. The effort is expected to sharply reduce repeated physical test drives and shorten development cycles. (Source: NVIDIA)

Phaidra's self-learning agents add 10% more compute within fixed data center power

Thu, 22 Oct 2026 05:00:25 GMT

10% more compute: Phaidra applied self-learning AI agents built on NVIDIA DSX Flex, DSX Max-Q and DSX Exchange to tune cooling and power allocation in data centers. The result is roughly 10% more usable compute under the same power budget, plus better grid flexibility. Emerald AI is also adopting the platform for power-constrained sites. (Source: NVIDIA)

Shopify cuts Sidekick serving costs 90% with post-trained NVIDIA Nemotron model

Thu, 22 Oct 2026 05:00:23 GMT

90% lower cost: Shopify post-trained an open NVIDIA Nemotron model to run the GraphQL subagent behind its Sidekick merchant assistant. The model matched a leading frontier model's accuracy on an internal benchmark while cutting serving costs 90% and responding seven times faster. Separately, Photoroom's image pipeline helps Vestiaire Collective standardize about 40,000 listing photos daily, with listings… (Source: NVIDIA)

AXA, Vaisala Xweather and TotalEnergies deploy NVIDIA Earth-2 weather models

Thu, 22 Oct 2026 05:00:19 GMT

1,000+ seasons: AXA is using NVIDIA's open Earth-2 weather models to simulate regional cyclone risk across more than 1,000 seasons. Vaisala Xweather is adding lightning data to its StormScope nowcasting tool for short-term warnings, and TotalEnergies is testing subhourly solar-radiation forecasts at two PV sites. No final quantified results have been reported yet. (Source: NVIDIA)

Poland's eHealth Centre builds AI imaging platform to speed urgent X-ray and CT case triage

Thu, 22 Oct 2026 05:00:02 GMT

Poland's eHealth Centre is building an Intelligent Services Platform on NVIDIA DGX BasePOD with the MONAI framework to help physicians analyze X-rays, CT and MRI scans and flag abnormalities. The tool is meant to prioritize urgent cases and cut the time patients wait for results. Separately, Poland's labor ministry, with Deloitte, is developing an agentic system for its 400-plus labor offices. (Source: NVIDIA)

Perplexity and NVIDIA researchers cut RL training tokens 6.5x with JOLT

Thu, 22 Oct 2026 04:56:31 GMT

6.5x fewer tokens: Researchers at Perplexity and NVIDIA introduced JOLT, a post-training method that trains one language model as both a privileged-information teacher and a student. On the MATH benchmark it matched GRPO's accuracy while processing 6.5 times fewer training tokens. A JOLT+ variant also lifted AppWorld pass rates to 72.09% from 62.20% for GRPO. (Source: MarkTechPost)

Google CEO Sundar Pichai uses AI agents to summarize daily updates and track compute usage

Thu, 22 Oct 2026 04:01:01 GMT

1 hour saved per day: Google CEO Sundar Pichai says he now relies on AI agents that condense the previous 24 hours of internal activity, follow threads across chat groups, and monitor compute usage by product area. He says the agents save him roughly an hour a day, while humans still make final resource-allocation decisions. (Source: Business Insider)

Meta's Muse AI app passes 10 million users, with shopping a leading use

Wed, 21 Oct 2026 21:14:19 GMT

10M+ users: Meta reported on its Q3 earnings call that its AI app Muse has more than 10 million users. Shopping and personal productivity are among the most common ways people use it, according to the company. (Source: TechCrunch)

Recursive releases Flint, a free laptop coding model largely built by its own AI research system

Wed, 21 Oct 2026 20:27:41 GMT

Startup Recursive launched Flint, a free coding model small enough to run on a laptop. The company says its AI research system did most of the work in building the model, serving as an early test of AI that improves itself. No performance figures were available in the excerpt. (Source: TechCrunch)

U.S. Hunger gives leaders direct access to organizational data with Amazon Quick

Wed, 21 Oct 2026 19:19:50 GMT

The nonprofit U.S. Hunger uses Amazon Quick, connected nightly to its SQL Server data via SPICE, so leaders can explore program and financial data without waiting on new data pulls. CEO Rick Whitted now follows his own questions through Full Cart food-assistance datasets, with shared definitions from technical staff. (Source: AWS Machine Learning)

OpenAI and Google's Wiz use AI to uncover long-hidden PHP security flaws before attackers do

Wed, 21 Oct 2026 17:59:12 GMT

AI tools from OpenAI and Google's Wiz were used to locate old, long-standing vulnerabilities in PHP, which were reported and addressed before exploitation. The article also cautions that the same AI capabilities raise risks if misused. (Source: TechCrunch)

Security researchers using OpenAI's model uncovered critical PHP flaws before attackers could exploit them

Wed, 21 Oct 2026 17:48:19 GMT

2 critical flaws found: Cybersecurity firm Calif used GPT-5.6 Sol to identify two critical vulnerabilities in PHP, the open-source language behind many websites. Wiz independently found the same issues, and OpenAI's "Patch the Planet" initiative helped drive fixes as governments urged updates. (Source: Business Insider)

Ineffable Intelligence and WPP test NVIDIA Vera Rubin instances on Google Cloud for AI workloads

Wed, 21 Oct 2026 10:45:17 GMT

Google Cloud's new A5X instances, built on NVIDIA Vera Rubin NVL72, are generally available next month after early access. In WPP's initial tests, creative agent workflows that normally needed a multi-node cluster ran on a single A5X instance. (Source: NVIDIA)

Nscale's tests show NVIDIA DSX MaxLPS fits 37% more GPUs in the same power budget

Wed, 21 Oct 2026 10:30:41 GMT

37% more GPUs: European AI cloud provider Nscale tested NVIDIA DSX MaxLPS on GB300 NVL72 systems to match power allocation to workloads. The tests showed 37% more GPUs and nearly 50% higher aggregate inference throughput within the same provisioned power. Nscale plans Vera Rubin deployments across five European markets. (Source: NVIDIA)

Siemens cuts circuit-board inspection calibration time 40% with synthetic AI data

Wed, 21 Oct 2026 10:30:40 GMT

40% faster: Siemens trained its circuit-board inspection system at an Erlangen electronics factory using synthetic defect images created with NVIDIA Cosmos Data Factory tools. The approach reduced inspection calibration time by 40%. Siemens also estimates 20% cost savings from factory agents that help identify root causes. 40% less inspection calibration time; est. 20% cost… (Source: NVIDIA)

Google Cloud releases llm-d 1.0 open-source inference stack for LLM serving fleets

Wed, 21 Oct 2026 04:33:40 GMT

85% lower median TTFT: Google Cloud and partners released llm-d 1.0, a stable Kubernetes-native layer that coordinates vLLM or SGLang model servers into one pool. A vendor-reported Ironwood TPU coding benchmark showed median time-to-first-token cut by more than 85% and 2.4x throughput. (Source: MarkTechPost)

Otter.ai raises $320M in funding driven by paying customers

Tue, 20 Oct 2026 22:24:09 GMT

$320M raised: Otter.ai, which makes an AI meeting assistant, secured $320 million entirely from General Catalyst's Customer Value Fund. That fund is tied to the revenue customers generate, signaling that paying users are fueling the company's growth. (Source: SiliconANGLE)

i14 trains and serves Laplace 3B linear attention model on AWS Trainium2

Tue, 20 Oct 2026 21:14:10 GMT

Fixed-size state: i14, working with AWS's GenAI Innovation Center, rebuilt its Laplace 3B linear attention model for AWS Trainium2 chips using native PyTorch via TorchNeuron Native. The design uses a fixed-size recurrent state, so serving memory stays flat as context length grows. (Source: AWS Machine Learning)

SpaceX's AI business drove its move into third-quarter profitability

Tue, 20 Oct 2026 20:40:41 GMT

$4.4B net income: SpaceX reported net income of $4.4 billion in the third quarter, reaching profitability. The company credits its AI business as the main driver of the result, which reportedly beat expectations. The source text is brief, so details of the AI products and revenue mix are not confirmed. (Source: TechCrunch)

Dell runs fully autonomous AI agents across its internal business operations

Tue, 20 Oct 2026 19:12:20 GMT

Dell Technologies has spent close to two years running fully autonomous AI agents inside its own operations. The company argues that reorganizing people and processes around these agents, rather than just adding tools, is what gives it an edge. The article gives few specific figures. (Source: SiliconANGLE)

Supply chain managers use Amazon Quick and AgentCore to replace a multi-tool morning routine

Tue, 20 Oct 2026 17:52:50 GMT

45-60 min routine: AWS describes a conversational workflow where supply chain managers ask Amazon Quick in plain language to spot disruptions, assess exposure, and draft an emergency RFQ. AgentCore-hosted MCP servers pull live data from ERP, monitoring, and procurement systems to replace a 45-60 minute routine across five tools. (Source: AWS Machine Learning)

AWS team fine-tunes GPT-OSS-20B to reason in four European languages

Tue, 20 Oct 2026 17:48:27 GMT

22% baseline: AWS engineers fine-tuned the open-weight GPT-OSS-20B model on Amazon SageMaker HyperPod using supervised fine-tuning, with GRPO added in some runs. The model's reasoning matched the requested language in French, German, Spanish and Italian far more often after training than the 22% baseline on a 50-example test. Multilingual reasoning in French, German, Spanish… (Source: AWS Machine Learning)

Octave Reliance builds tenant-isolated AI foundation on Amazon Bedrock

Tue, 20 Oct 2026 17:34:04 GMT

50% faster deploys: Octave Reliance, a quality management SaaS for regulated industries, built a multi-tenant AI foundation on Amazon Bedrock using Cohere Embed v4. The platform embedded roughly 2 billion documents and launched a Complaint Triage Assistant. The shared foundation halved deployment time for each new AI use case. (Source: AWS Machine Learning)

Meta's Muse AI app reaches 10 million US downloads in 40 days

Tue, 20 Oct 2026 17:12:48 GMT

10M downloads: Meta's AI app Muse crossed 10 million US downloads within 40 days of launch, a pace matched by only 11 other apps. Early data also suggests users stick with Muse at higher rates than rival AI apps. (Source: TechCrunch)

Eon consolidates multi-cloud investigation tools behind one AI agent gateway

Tue, 20 Oct 2026 15:57:57 GMT

Eon, a cloud data platform vendor, used Amazon Bedrock AgentCore to build an MCP Gateway that replaces per-tool credentials and local MCP processes. Its engineers and autonomous SRE agent now reach investigation tools across AWS, GCP, and Azure through one authenticated entry point. (Source: AWS Machine Learning)

ALTK-Evolve learned memory tops τ²-bench customer service leaderboard

Tue, 20 Oct 2026 15:19:48 GMT

#1 on τ²-bench: Researchers reported that ALTK-Evolve, a learned-memory system for AI agents, reached the top of the τ²-bench customer service leaderboard. The result came without depending on a frontier-class model, pointing to memory design as a lever for agent performance. (Source: Hugging Face)

Spotify removed over 50 million AI spam tracks from its recommendations

Tue, 20 Oct 2026 14:18:37 GMT

50M+ tracks removed: Spotify says it used its recommendation system to filter out more than 50 million tracks it classed as AI-generated slop or spam over the past year. It also shifted recommendations toward human-made music and reports a marked drop in listener complaints about such content. (Source: The Verge)

Cloudflare cascade of AI models cut damaged-product claim processing cost 49%

Tue, 20 Oct 2026 07:45:41 GMT

49% lower cost: Cloudflare tested its Clairic extraction model on Workers AI in a damaged-product claims pipeline, chaining it with its Clef decision model and GLM-5.3 for drafting replies. The cascade ran about 3.2x faster than GLM-5.3 alone and lowered cost per million claims by 49%, though Clairic still struggles with multi-step arithmetic. (Source: MarkTechPost)

Three Spanish startups deploy AI in water, cardiology and pharma sales

Tue, 20 Oct 2026 05:00:00 GMT

Three startups from the Barcelona Supercomputing Center's AI Factory Accelerator Programme have moved AI into production or deployment. Qatium uses its water-network AI in early production, OneCareAI built a tool using NVIDIA NIM to flag hidden atrial fibrillation on ECGs, and Avatia launched a voice-first multi-agent system for pharma sales teams. (Source: NVIDIA)

Deepgram optimizes voice AI models at scale on AWS Trainium chips

Mon, 19 Oct 2026 16:23:22 GMT

1,400 organizations: Deepgram, which builds speech-to-text, text-to-speech, and voice agent models, moved its workloads to AWS Trainium to run its voice AI at scale. The company now handles quadrillions of audio tokens a year for more than 200,000 developers and 1,400 organizations. (Source: AWS Machine Learning)

Adobe Frame.io speeds video search for UFA with NVIDIA GPU-accelerated AI pipeline

Mon, 19 Oct 2026 15:00:31 GMT

Adobe is using NVIDIA accelerated computing in Frame.io to help creative teams search video content faster. The GPU-accelerated pipeline combines NVDEC decoding, TensorRT inference and CUDA preprocessing, and is helping German producer UFA locate the right moments more quickly. (Source: NVIDIA)

Swing Left's Ground Truth uses LLM-powered voice-to-text app to mine voter conversations

Mon, 19 Oct 2026 10:31:00 GMT

3,000+ volunteers: Swing Left's Ground Truth project has sent more than 3,000 volunteers to over 50 House districts in 23 states to hold open-ended "deep canvassing" conversations with voters of all leanings. Canvassers log voice-to-text recaps in an AI app, and organizers query the stored records in plain English to surface findings for campaigns. (Source: Fast Company)

Idealab's Bill Gross builds 1.2 million AI-generated business ideas database

Mon, 19 Oct 2026 10:00:00 GMT

1.2M ideas: Incubator founder Bill Gross fed about 7,000 news articles into Anthropic's Claude, which produced 32,000 business ideas in an initial test. He then widened the inputs to patents, SEC filings and consumer complaints, building a database of over 1.2 million ideas that Idealab.ai plans to offer to aspiring founders. (Source: Fast Company)

Microsoft moves Maia 200 AI chips into production for Copilot and Teams

Mon, 12 Oct 2026 13:01:29 GMT

1.7x tokens per watt: Microsoft says its in-house Maia 200 accelerator is now running in production across the US, Europe and Australia, serving Copilot and built-in AI features in Teams. The company reports up to 1.7x more tokens per watt than leading GPU infrastructure and plans to make Maia its main inference system from 2027. (Source: Microsoft)