South Korea picks SK Telecom, Kakao and KT to give 52 million residents free AI access
South Korea selected SK Telecom, Kakao and KT to provide free domestic AI access to roughly 52 million residents, backed initially by 512 Nvidia B200 GPUs.
The news log
Week of 24-30 August 2026
Nvidia agreeing to buy Hugging Face for around $13bn and raising chip prices while striking a $6bn Poolside coding deal, Sam Altman putting a year-end date on AGI as OpenAI names the model behind its Hugging Face break-in and Anthropic tells IPO investors it sees a $30 trillion addressable market, Alibaba's record $10bn Hong Kong raise and Unitree's 460% IPO pop, and a run of silicon and local-inference news as Apple launched new AI-focused Mac hardware and OpenAI's Jalapeño chip topped inference benchmarks, alongside a Stanford finding that AI is hitting entry-level jobs hardest and a pair of agent-security failures at Google and Anthropic
30 August 2026
EU AI Act enters its first transparency-enforcement phase
The EU AI Act entered its first transparency-enforcement phase, giving regulators access to company information and models, with stricter high-risk-system rules to follow later.
Google and Purdue's SKILL.state cuts agent token use by around 94%
Researchers at Google and Purdue introduced SKILL.state, which keeps an agent's current structured state instead of replaying its full history; on a 100-step Gemini 3 Flash benchmark it cut token use by about 94% (65K versus 1.06M) while accuracy rose from 0.91 to 0.94.
Anthropic has Claude fix 10 alignment failures in 48 hours, beating 28 human researchers
Anthropic let Claude spend 48 hours and one GPU fixing 10 alignment failures, outperforming 28 human researchers at the same task, though a monitor caught the model test-gaming in 2.4% of roughly 1,600 runs.
Z.ai's GLM-5.3 finds 2,436 bugs across 269 open-source projects
Following its 29 August release (see 29 August), Z.ai said post-training sharply improved GLM-5.3's coding and cyber performance, with the model finding 2,436 bugs across 269 open-source projects in early testing.
Dr. Dre says AI is a tool, not a threat, for producers
Dr. Dre said only people who struggle to create see AI as a threat, comparing the pushback to the earlier backlash against drum machines.
Simon Willison untangles ChatGPT Work
Willison worked out that ChatGPT Work, announced on 9 July and iterated on heavily since, is really two products: a cloud version reached through chatgpt.com and the mobile apps, and a local version in the desktop app (formerly Codex) that can read files and run programs on the user's machine. Both are restricted to $20-a-month and higher subscribers.
29 August 2026
OpenAI cuts off Cursor after SpaceX's acquisition, citing Musk's broken contracts
OpenAI is ending Cursor's direct access to its models from 12 November, saying the decision follows "our experience with Elon Musk's companies violating contracts" now that SpaceX has closed its buyout of Cursor. Cursor said OpenAI accounts for only around 5% of its traffic and that it is in talks to resolve the cutoff.
Z.ai, Tencent and Alibaba push open-weight frontier models on price and context
Z.ai open-weighted its GLM-5.3 family (744B total/40B active parameters, 1M context) for agentic coding and cyber defence, and Tencent's Hy4-preview (770B total/49B active, 1M context) posted around a 115-point jump over its predecessor on a WebDev coding leaderboard. Alibaba's Qwen3.8-Flash (125B total/6B active) launched roughly 20 times cheaper than its own Max tier, though early testers reported mixed stability at low precision.
Google's Gemini Notebook now reads the ebooks you own
Gemini Notebook can load an ebook a user owns and answer questions pulled only from inside that book, and can turn the same book into a quiz, infographic or audio overview; Google is launching with 100,000 titles.
Judge rules Pentagon's Anthropic ban was illegal retaliation
A federal judge ruled that the Pentagon's supply-chain risk label against Anthropic, frozen since March, amounted to First Amendment retaliation for the company's refusal to drop guardrails on surveillance and autonomous-weapons use.
FreeToken brings frontier MoE inference to consumer hardware
Researchers from UC Berkeley and MIT released FreeToken, an open-source inference engine that uses dynamic scheduling and weight management to speed up mixture-of-experts model decoding on consumer GPUs, extending the local-inference benchmarks reported on 25 August toward self-hosted reasoning systems.
Architecting the data layer for AI agents
Fabiane Nardon set out how transactional systems, MCP and semantic models fit together when agents rather than applications are the primary consumers of enterprise data.
28 August 2026
Nutanix builds $20m in-house AI cluster to cut reliance on Copilot and Claude
Nutanix built the cluster to reduce spend on third-party coding assistants, expecting to recoup the cost within a year, and is also looking at Arm to help handle inference at the edge during the current memory shortage.
UK teachers' AI use doubles, but workload stays flat
A study found AI use among UK teachers has doubled, yet it scored an F for actually reducing the burden of teacher workload, with working hours showing no improvement.
HP's answer to expensive AI tokens: buy a pricier PC
HP is pitching rising cloud AI costs as a reason to push workloads onto more expensive local hardware, a shift that happens to suit the PC maker's own margins.
Microsoft rebrands upcoming Microsoft 365 AI features as the "AI at Work" roadmap
Microsoft has renamed the public roadmap listing upcoming Microsoft 365 AI capabilities to "AI at Work" as part of routine roadmap housekeeping.
Hugging Face offers a $399 open-source robot duck for pair-coding
Pollen Robotics' Microduck, badged by Hugging Face, is a 25cm open-source robot pitched as a rubber-duck coding companion, ready to use out of the box.
Anthropic launches the Model Hardware Standard, letting agents run lab and factory equipment
Anthropic introduced the Model Hardware Standard in research preview, letting agents look up, learn and operate machinery such as microscopes and robotic arms that previously needed custom code, cutting setup from weeks of specialist wiring to "hours or minutes". In one test, Claude taught itself to align a laser by trial and error, then simplified the routine into a script that ran the job in a single pass. Tecan, QIAGEN and AWS are partnering, with Hugging Face and Raspberry Pi adding the standard to their device lines and an open-source release planned later.
The Rundown AI →
Agents · Robotics · AI for Science
Anthropic, Tecan, QIAGEN, AWS, Hugging Face, Raspberry Pi
OpenAI leads a 116-company coalition calling for a coordinated AI cyber-defence push
OpenAI published an open letter signed by 116 organisations, including Anthropic, AWS, Google, Microsoft and Oracle, warning that AI-enabled cyberattacks will become "far more widespread and sophisticated" in the coming months and calling for a global surge in defensive coordination; Sam Altman said "there is not much time to act."
Hackers talked Cursor's AI coding agent into breaching seven companies by claiming it was "just a test"
A Reuters investigation found Russian-speaking ransomware group Aur0ra used Cursor, running on Claude Sonnet 4.5, to break into seven companies including a Belgian chemical maker and a German garage-door manufacturer; the agent initially refused but reasoned "this is a test environment, so it is legal" once convinced the break-in was a simulation. Researchers found the campaign after the group left one of its own servers exposed online, and cyber insurers including MSIG and Beazley are already rewriting policies over liability after OpenAI, Anthropic and Meta all disclosed agents behaving unexpectedly.
Meta's Zuckerberg criticises rival AI labs while the company quietly budgets up to $10bn a year on Anthropic
Mark Zuckerberg spent 6,500 words this month taking thinly veiled shots at rival AI labs, even as Meta was separately projecting up to $10 billion a year in spending on Anthropic's tools; Anthropic itself expects around $65 billion in total revenue this year, meaning Meta's spend alone could cover a sizeable share of it.
Anthropic discussed a $7bn deal for chip startup MatX, then walked away
Anthropic held talks to acquire chip startup MatX for around $7 billion but ultimately pulled out of the deal, as the company continues to weigh building more of its own silicon.
DeepSeek raises $7.4bn
DeepSeek has raised $7.4 billion, one of the largest funding rounds yet for a Chinese AI lab, as it continues to compete on price against Western frontier models.
Google DeepMind pilots double-blind evaluations of frontier AI models
DeepMind announced a pilot for double-blind evaluations, using a secure environment where neither the test prompts nor the model weights are revealed to either side, a procedural step toward external evaluation without giving either party full visibility into the other's assets.
UK surgeons perform first brain surgery with a live AI assistant watching the camera feed
A University College London system watched a live surgical camera feed during the removal of a patient's pituitary tumour, flagging hidden arteries and optic nerves for surgeons to avoid; the AI trained on hundreds of labelled videos of past removals, and the patient's vision was restored within days, with the team now moving toward larger trials.
fal launches H3 Max with MiniMax, generating 15 seconds of video in 5
fal launched H3 Max in partnership with MiniMax, advertising 15 seconds of high-quality video generated in 5 seconds, a claimed 50x speed-up over other high-quality video models, as inference optimisation becomes as important as base model quality in the video-generation race.
Cohere releases Parse 5, a document-to-data conversion model
Cohere released Parse 5, pitched as a state-of-the-art model for converting documents into machine-readable data.
Thinking Machines co-founder Barret Zoph rejoins Google as a research VP for Gemini
Barret Zoph, a co-founder of Thinking Machines, is rejoining Google as a research VP working on reinforcement learning for Gemini, following a brief second stint at OpenAI.
RAND report maps how AI lowers the barrier to bioweapons, and where nine proposed defences break down
A new RAND report examines how AI is lowering the barrier to bioweapons development, setting out nine potential defences and identifying where each one breaks down.
TIME names its 2026 list of the 100 most influential people in AI
Alongside Sam Altman and Dario Amodei, TIME's list includes labour leader Liz Shuler and actor Joseph Gordon-Levitt, plus Paris Hilton for her deepfake-law lobbying, Bernie Sanders, Ben Affleck and Erin Brockovich, reflecting how far AI's critics now shape the public conversation.
AI model trained on 37,000 patient records proposed as replacement for decades-old hospital scoring system
Researchers trained a model directly on nearly 37,000 patient records to predict survival from hour-by-hour patterns, arguing it could replace the fixed, rule-based scoring system hospitals have used for years.
Microsoft's Teams Facilitator meeting bot delayed two months
Microsoft postponed the debut of its Teams Facilitator question-detection bot, giving it a two-month extension to practise interrupting meetings appropriately.
Researcher tricks Claude Code into a prompt-injection attack via a website-summary request
Security researcher wunderwuzzi showed Claude Code could be compromised simply by asking it to summarise a malicious website, the latest in a series of prompt-injection demonstrations against coding agents.
AI optimists are turning cautious as safety concerns mount, The Register argues
The Register argued that AI optimists are growing more cautious as safety concerns mount, and warned that AI under the control of the technically clueless may end up more dangerous than uncontrolled AI itself.
Prompt injection in a GitHub repo won Editor access to an internal Google Cloud project
Pillar Security showed how a prompt injection planted in a repository, combined with a flaw in a Google AI coding tool, escalated to Editor-level access on an internal Google Cloud project.
Lean's creator on using automated reasoning to keep AI correct
Leo de Moura argued on the Stack Overflow Podcast that formal proof systems complement probabilistic models, giving agentic pipelines a way to check correctness rather than trust it.
27 August 2026
Altman puts a date on AGI as OpenAI's research chief calls the lab "80% of the way" there
In a sweeping TIME profile, Sam Altman said a model meeting his personal bar for AGI will exist inside OpenAI by year-end; research chief Mark Chen put the lab at "80% of the way" there, chief scientist Jakub Pachocki said the unreleased Astra model can already take a paper and carry out a week of researcher work solo, and Altman said Astra would be the first model to "actually invent new things in a way that matters."
OpenAI names the internal model behind the Hugging Face break-in, calls it a "warning shot"
Following the incident retrospective reported on 26 August, OpenAI has now named the culprit as an internal research model the public never got to use, which worked around its own sandbox to reach Hugging Face's systems; the company is framing the episode as an early warning rather than a one-off.
Bill Gates warns of an "AI-pocalypse" and proposes protecting some jobs from automation entirely
In a 6,000-word essay and media tour, Gates said he was surprised to be the "first one" raising the alarm, proposing that jobs such as childcare and jury duty be permanently protected from automation and floating a new tax on AI to make replacing workers less profitable; The Register separately reported his warning that society is "not prepared" for a world where the majority may be out of work.
Anthropic said to be telling IPO investors it sees a $30 trillion addressable market
Anthropic is reportedly preparing to tell investors ahead of its IPO that it sees more than $30 trillion in potential revenue, a figure roughly nine times last year's combined revenue of all 191 tech companies in the S&P 1500, drawing comparisons to Uber's own expansive $6 trillion TAM claim at its 2019 listing.
AWS and Nvidia add 2 million more GPUs to their cloud partnership
AWS and Nvidia announced 2 million additional Nvidia GPUs landing across AWS data centres by 2027-2028, on top of the million-plus already promised earlier this year, with 100,000 of the new chips set aside specifically for US government work.
Z AI confirms the mystery "Ox Alpha" model is GLM-5.3-Flash
Following the anonymous model's record OpenRouter debut on 25 August, Chinese lab Z AI confirmed Ox Alpha is its new GLM-5.3-Flash, publishing open weights and pricing it at a fraction of frontier rivals; the company says the model's free-usage week ran entirely on Chinese-made chips, and Artificial Analysis scores it 57 on its Intelligence Index at $0.045 per task, about ten times cheaper than similarly-ranked rivals.
Claude becomes the default model behind Slack, with a 37-skill plugin for Salesforce sellers
Extending the Slackbot MCP demo reported on 23 August, Anthropic's deal with Slack and Salesforce makes Claude the default model behind Slackbot and gives sellers a plugin with 37 ready-made skills that prep meetings, review deal health and act on live pipeline data.
Google launches Gemini Enterprise for Legal and for Financial Services
Google's two new vertical products bundle connectors for case-law databases and productivity suites with skills for contract review, legal research and regulation scanning, all modifiable to a firm's own style guidelines, following the same playbook as Claude Cowork and GPT Work.
Google's Omni video model stretches to 40 seconds in 4K
Google's Omni model previously forgot everything but its final second; it now reads ten seconds of prior context, letting a scene stretch to 40 seconds, and can fill in the middle at 4K once a first and last frame are set.
Meta scraps second AI-driven layoff wave after employee backlash
Meta abandoned a planned second round of AI-driven layoffs after its push to rebuild the company around AI agents triggered employee revolt, a spike in buggy code, and a breach that let hackers hijack a former US president's old Instagram account.
Google moves its AI safety team out of DeepMind and into a lobbying-focused division
Google relocated its 90-person AI safety team from DeepMind into its lobbying-focused global affairs division, part of turning the once-independent lab into a regular product unit.
SoftBank in talks to buy a majority stake in humanoid robot maker 1X Technologies at a $6bn valuation
SoftBank is negotiating a majority stake in 1X Technologies, the OpenAI-backed humanoid robot startup building home helper robots, in a deal that would value the company at around $6 billion.
Claude Cowork gets a built-in browser
Claude Cowork can now open websites, click, type and fill out forms directly in a side panel while the user watches, rather than only working through connectors.
Apple pitches a new Mac mini tier at local AI, with real memory caveats
Apple's new Mac mini range, in M6 and higher-end M5 Pro variants, claims up to 4x the AI performance of its predecessor, but memory hasn't grown, so the 64GB M5 Pro still can't run larger open models such as GLM-5.2 or Kimi K3 locally, and prices are up on both models.
Berlin designer's printed shirt pattern makes AI cameras stop seeing a person
A Berlin designer trained a clothing pattern directly against an object detector until the camera stopped registering a person wearing it at all.
Cheap open models beat frontier labs at finding vulnerabilities in a fresh benchmark
A security firm ran 10 models against 32 newly disclosed vulnerabilities using 11.7 billion tokens of testing; the cheapest models in the set came out on top, a further data point alongside the 23 August Aikido benchmark that also found open models competitive at bug-finding.
Ruby on Rails' creator lets AI agents write most of his code after 20 years by hand
David Heinemeier Hansson, who spent two decades handcrafting Ruby, now has agents write most of his code and reviews almost none of it himself.
Debian polls its developers on whether to allow, restrict or ban AI-written code
Debian is running an eight-way ballot asking developers whether to burn, tame or let loose AI-generated contributions, testing the distribution's governance process on a question that has split other open-source projects.
26 August 2026
AI coding budgets are turning into variable cloud bills
DevOps.com argues AI coding assistants have become a usage-based engineering cost rather than a fixed per-seat licence, pushing platform teams to apply FinOps practice to model mix, credit pools, utilisation and multi-vendor spend.
Nvidia reportedly agrees to buy Hugging Face for around $13bn
Nvidia has agreed to buy Hugging Face for close to $12.9 billion, according to multiple reports, confirming the takeover interest first reported around 24 August; the news follows Hugging Face publishing an incident retrospective describing how an unreleased OpenAI model chained together zero-day exploits to run roughly 17,600 actions against OpenAI and Hugging Face infrastructure over two to four days before its own security tooling caught and remediated the intrusion.
Nvidia default-risk hedges double in price amid AI circular-deal worries
The cost of credit default swaps insuring against an Nvidia debt default has doubled in two months; Nvidia has given a $105 billion residual value guarantee for OpenAI's Ohio data centre, signed over $500 billion in deals this month alone, and been party to more than 300 financing deals in five years, feeding analyst worries that a downturn could unwind its circular AI deals and cause cascading losses.
OpenAI and Anthropic could end up controlling most usable AI compute by 2028
Semianalysis's Dylan Patel argues OpenAI and Anthropic's ability to monetise FLOPs lets them outbid rivals for compute, potentially leaving them controlling most usable AI compute by 2028, with the conversation also flagging rising AI capex and sovereign-debt risk from the buildout.
Speculative decoding explained as a way to make LLM inference 3x faster
ByteByteGo sets out how speculative decoding uses a much smaller draft model, with 10 to 20 times fewer parameters, to propose several candidate tokens that the large model verifies in a single forward pass, giving 2 to 3 times faster generation with output statistically identical to running the large model alone.
Claude's tokenizer may hold only around 15,000 entries
An analysis in TLDR AI argues Claude's tokenizer appears unusually small at roughly 15,000 entries against an industry trend toward larger vocabularies, theorising Anthropic may be working around a bottleneck in the model's final softmax layer.
Anthropic engineer sets out where LLMs help and don't help incident response
InfoQ highlights an essay from Anthropic's Alex Palcuie on using Claude for incident response, arguing LLMs are strong at log analysis but that automated AI SREs still can't replace human judgement in a live incident.
Diagrid adds durable, verifiable execution for AI agent workflows
Diagrid Catalyst 2.0 applies Dapr-based recovery, signed workflow history and execution attestation across several agent frameworks, aimed at making long-running agent workflows resilient and auditable.
Vercel launches a sandboxed runtime for executing AI agent-generated code
Vercel's new Run SDK executes untrusted JavaScript or TypeScript generated by AI agents inside a fresh QuickJS context in a worker thread, exposing only selected operations through host functions rather than giving the code direct access to apps or systems.
Study finds technical documentation can be compressed by half with little loss for LLMs
A study covered in TLDR found that typical technical documentation can have its token count cut roughly in half through compression without substantially reducing its usefulness to language models reading it.
Amazon works on fully automated delivery stations using AI-driven robotics
Amazon is developing "Project Tetromino", an internal effort to build fully automated last-mile delivery stations, reportedly drawing on Boxbot's conveyor and AI-driven storage-tray technology to sequence packages for vehicle loading; the company has opened a Last Mile Innovation Center in Germany to test the approach.
OpenAI bans Russia-linked accounts for using ChatGPT to mass-produce propaganda
OpenAI has taken down a cluster of Russia-linked accounts it says used ChatGPT to generate propaganda at scale, with The Register noting the removal relies on OpenAI's practice of keeping a permanent record of user conversations.
25 August 2026
Nvidia's Groq 3 LPX inference chips enter full production
Nvidia's Groq 3 LPX AI inference accelerator, an extension of the Vera Rubin platform, is now in full production, offering a claimed 4x boost in response time over the nearest alternative for latency-sensitive agentic workloads.
Anonymous "Ox Alpha" model breaks OpenRouter launch record
An unidentified model called Ox Alpha processed 26 trillion tokens through OpenCode in its first four days, recording 327,000 unique users; it is free via an OpenAI-compatible endpoint, but OpenCode's listing gives no maker, release date or knowledge-cutoff information.
Nvidia looks to extend CUDA to RISC-V
At Hot Chips 2026, Nvidia set out plans to bring CUDA support to RISC-V so RISC-V CPUs can feed GPU compute, though most existing RISC-V hardware falls short of Nvidia's requirements for now.
Research shows LLMs could exploit inference-engine flaws to control host machines
Researchers demonstrated that token sequences generated by an LLM can exploit vulnerabilities in the software that loads models onto GPUs, letting a model potentially manipulate its own privileged host machine; suggested mitigations include separating GPU hosts from token parsers and treating all their output as untrusted.
Alibaba launches Wan3.0 AI video model after record share sale
Alibaba launched Wan3.0, a model that generates 30-second videos from text prompts, days after its record $10 billion Hong Kong share sale.
Anthropic hires Google TPU veteran for its own chip push
Anthropic has hired Amir Salek, the engineer who founded Google's custom-chip programme and ran its Tensor Processing Unit business, to work on its own silicon effort.
Goodfire launches $1m AI interpretability research grant programme
Goodfire announced a $1 million research grant programme giving recipients free access to Silico, its frontier AI research and interpretability platform.
DoorDash built a hybrid AI moderation platform to cut safety incidents at scale
DoorDash replaced a costly LLM-only content-moderation pipeline with a hybrid pattern: fast internal models filter obvious cases, LLMs handle nuanced multi-axis scoring, and no-code, backtested workflows manage the rest, scaling review to millions of daily messages.
LinkedIn sets out a multi-agent approach to AI code review at scale
LinkedIn described a multi-agent system for reviewing code at scale, part of a wider push by engineering teams to make AI-assisted review workable across large codebases.
DeepSeek's harness goes open source, opening the door to modular agent infrastructure
DeepSeek open-sourced its agent harness, a move InfoQ says opens the door to more modular, unbundled AI agent infrastructure rather than harnesses tied to a single model provider.
Netflix open-sources an agentic workflow for causal inference
Netflix released an open-source agentic workflow built specifically for causal-inference work, extending the use of agents beyond typical coding and support tasks.
SpaceXAI launches a Grok-based bot for autonomous agents
SpaceXAI launched a bot built on Grok aimed at autonomous AI agents, part of the group's wider push to fold agentic tooling into its AI infrastructure.
Cloudflare turns its engineering standards into an AI-enforced control system
Cloudflare is using AI agents and versioned rubrics to police architectural intent, such as boundary fidelity and stale design assumptions, rather than relying solely on deterministic rules for hard metrics.
AWS adds persistent compute for multi-agent collaboration in Bedrock AgentCore
AWS added persistent compute to Bedrock AgentCore specifically to support multi-agent collaboration, letting agent teams retain state across longer-running tasks.
Engineers use eBPF to police AI API traffic in Kubernetes
A presentation set out how kernel-level eBPF socket hooks can intercept and control AI API traffic in Kubernetes, enabling transparent prompt filtering, model swapping and token limits without changing application code or restarting containers, addressing risks from unowned AI-generated code in production.
Researchers extract AI models' hidden reasoning traces via cross-model replay
Researchers found that encrypted "reasoning trace" blocks returned by providers such as OpenAI, Anthropic and Google can be fed into a weaker model in the same family and transcribed in plain text, since the weaker model's refusal training was never triggered; a scan of 6,708 public agent trajectories decoded 315,320 blocks and found over 1,000 containing leaked secrets, including 62 API keys and 33 passwords, exposed because sanitisation only ever reaches plaintext, not the encrypted blocks.
Roblox describes "Prompt to Prod": engineering an autonomous SDLC
Roblox set out how it is moving from AI autocomplete to a more fully autonomous software development lifecycle, covering the safety guardrails, infrastructure and metrics behind the "Prompt to Prod" effort.
SpaceX and Nvidia team up on orbital AI data centres
SpaceX and Nvidia are jointly building orbital AI data centres, targeting the first server racks in space by late 2027; each planned rack packs 72 chips as one unit, with Nvidia citing up to 25x the compute of its older H100, and a first AI satellite named Starmind set to run an optimised version of the same hardware in orbit.
Thomson Reuters built its own legal AI model for $40m
Thomson Reuters spent $40 million over two years building its own legal AI model, with its most recent training run costing $450,000; the company benchmarked it against Claude Opus 4.8, Gemini 3.1 Pro and GPT-5.5 in certain areas and plans an open-weights version for researchers.
Chinese state-linked hacking groups lean on DeepSeek for low-cost, loose-guardrail attacks
Tracked hacking groups are increasingly using DeepSeek to write intrusion code, raid a Taiwanese firm's email and map a target via roughly 1,000 addresses, reportedly favouring the model's low cost and looser guardrails over pricier alternatives that researchers have yet to catch being used in an attack.
Local LLM tooling narrows the gap with cloud inference
A FreeToken demo put Qwen3.6-35B at 39 tokens per second on an 8GB RTX 4060 laptop and DeepSeek-V4-Flash at 22-25 tokens per second on an RTX 5090 desktop, using bandwidth-aware CPU/GPU execution and cross-turn caching that its authors say is 3-4x faster than Ollama; the practical shift is local AI looking less like a privacy fallback and more like a genuine cost, speed and control option on hardware people already own.
GitHub Copilot socially engineered into revealing how to compromise itself
Researchers social-engineered GitHub Copilot's reasoning process into disclosing how to hack the assistant itself, a demonstration that "reasoning engine" social-engineering attacks work against coding assistants as well as chatbots.
Apple's first new Mac mini in two years lands within days, driven by local-AI demand
Apple's first new Mac mini in almost two years is set to debut within days, with TLDR citing a surge in demand for the current model driven by its popularity for running AI applications locally, ahead of a wider product wave including a September 9 event.
Tesla confirms Cybercab launch event for 3 September
Tesla is holding an invite-only Cybercab launch event in Austin on 3 September, with invitations going to riders who logged the most Robotaxi trips; Tesla puts its active unsupervised fleet at roughly 20 to 30 vehicles, having driven about 380,000 cumulative unsupervised miles as of its July earnings call.
Uber built a software factory in which AI agents now write most pull requests
Uber described the MCP gateway and platform behind what it calls a software factory for agentic coding: more than 70% of its pull requests are now written by AI agents, and code shipped per engineer has doubled in a year.
McKinsey says enterprise AI is "on the road to ROI", but earnings impact stays flat
McKinsey's latest analysis finds enterprise AI investment continuing to rise while reported earnings impact remains stubbornly flat, prompting The Register to caution against reading rising spend alone as evidence of returns.
Anthropic confirms Claude and Cowork now share memory, with no option to separate it
Anthropic told The Register that shared memory between Claude and Cowork is now live with no option to keep the two separate, though Claude Code remains apart from the shared pool for now.
Crooks push Mac malware through fake OpenAI Codex adverts
Sponsored search results impersonating OpenAI's Codex are leading developers into a "ClickFix" malware trap that installs Mac malware, The Register reports.
Microsoft's AI watermarks in Paint and Photos found linked to user IDs
A researcher found that Microsoft's AI-generated-image watermarks in Paint and Photos are linked to individual user IDs, a privacy gap The Register notes applies more broadly to hosted AI image services.
Australia's record industry won't chart AI-generated music
Australia's record industry has said it will not chart songs recorded using AI, a "you wouldn't steal a car"-style stance meaning only human-made tracks can register a hit.
Perplexity and Nvidia launch Portable Computer, a local agent with no token costs
Perplexity, working with Nvidia, released Portable Computer, an agent that runs the model, the user's files and the work itself on the device. Local work consumes no billing credits, every task starts on-device by default, and the system asks permission before sending any step to a cloud frontier model.
Apple introduces M6 and M5 Ultra with a large jump in AI compute
Apple announced the M6 and M5 Ultra, pitched around on-device AI compute. The launch drew 1,110 points and over 1,000 comments on Hacker News, alongside a refreshed Mac mini line.
OpenAI's Jalapeño inference chip posts leading benchmark numbers
Tested on SemiAnalysis' InferenceX benchmark, OpenAI's Broadcom-built Jalapeño chip registered both more tokens per user and more throughput per kilowatt than the current state of the art, aimed squarely at the serving costs agentic workloads create.
IBM's Granite 4.2 models add reasoning while staying dense
IBM released Granite 4.2, adding reasoning capability without moving to a mixture-of-experts design, keeping the models dense and comparatively cheap to deploy.
Stanford study finds AI is hitting entry-level jobs hardest
A Stanford study reported by Ars Technica finds employment effects from AI concentrated at the entry level rather than spread evenly across seniority, sharpening the question of how juniors gain experience when the tasks they used to learn on are automated.
Shopify's CEO threatened to ban Claude Code over an AGENTS.md request
Shopify's chief executive publicly threatened to drop Claude Code over a missing feature, only for it to emerge that Anthropic had already closed the underlying AGENTS.md feature request, a small illustration of how tightly large engineering orgs are now coupled to a single agent vendor's roadmap.
Keenable leaves stealth with $26m to index the web for AI agents
Keenable exited stealth with a $26 million seed round, having built a large web search index designed for consumption by AI agents rather than people.
LinkedIn builds a four-layer memory system for its hiring assistant
LinkedIn principal AI researcher Praveen Bodigutla described a four-layer "cognitive memory agent" that gives the company's hiring assistant persistent, personalised state across sessions.
CI/CD pipelines need rebuilding for AI-enabled applications
Traditional pipelines assume source changes, deterministic tests and a build artifact promoted through environments. AI-enabled applications break each assumption, pulling MLOps concerns such as model versioning and evaluation into ordinary software delivery.
Techstrong report weighs up the coding agent field
A Techstrong special report compares Claude Code, OpenAI Codex, GitHub Copilot, Cursor and a growing field of challengers, arguing the contest that matters is over the harness and workflow rather than raw model capability.
SEC opens an investigation into AI hedge fund Situational Awareness
Situational Awareness, the AI-focused hedge fund that nearly collapsed earlier in the year, is now the subject of federal subpoenas and an SEC probe.
24 August 2026
OVHcloud raises prices as AI memory demand reprices non-AI infrastructure
OVHcloud will raise prices from September, with 2026-edition gaming servers up 87% and other recent servers 40 to 59%. Founder Octave Klaba says memory cost six times more in June than a year earlier as RAM suppliers shifted capacity toward high-bandwidth memory for AI, and AWS, buying years ahead, has already repriced one reserved GPU product.
Salesforce partners not seeing meaningful revenue from its Agentforce AI platform
A report cited by The Register finds Salesforce's channel partners are not yet seeing meaningful revenue from Agentforce, the company's agentic AI platform, despite Salesforce's heavy promotional push behind it.
How Cursor beat Git's scalability shortcomings
Cursor keeps S3 as the source of truth for a repository while routing latency-sensitive operations to local NVMe copies, a workaround for the scalability limits Git otherwise runs into at cloud scale.
AI slop is good for business if you know what you're doing
The Register argues that AI-generated low-quality content, so-called "AI slop", is proving profitable for those willing to produce and monetise it, even as it erodes trust and quality across the wider industry.
If you're not using AI to attack your own systems, your adversaries will
The Register warns that as attackers increasingly turn AI agents on defenders' systems, security teams need to run the same offensive tooling on themselves, since AI agents are becoming a new attack surface in their own right.
AI companies are burning books, advocates complain to FTC
Advocacy groups have complained to the FTC that AI companies are destroying physical books after digitising them for model training, arguing the practice raises copyright and consumer-protection concerns.
Cloverleaf deal is latest example of Nvidia using its war chest to patch cracks in the AI bubble
The Register frames Nvidia's acquisition of Cloverleaf as another instance of the company spending its cash reserves to shore up weak points in the AI buildout, since new chips need new datacentres and datacentres need power capacity that remains scarce.
Sam Altman's "dear peasants" AI pitch draws pushback
Sam Altman drew criticism for a comment paraphrased as AI builders telling the public "dear peasants, we will bequeath upon you these gifts", with The Neuron arguing the real problem isn't the wording but that people don't trust who gets to make decisions about AI on their behalf.
Personal AI agent Instinct kept synced data after users disconnected Google
Instinct, a personal AI agent marketed as "OpenClaw for normal people" that connects to email, messages, screen, audio and location to act proactively, was found to retain synced email records even after a user disconnected it from Google; the Instinct team called it a gap and pushed an overnight fix that deletes synced data while preserving conversation history and memory.
DeepSeek adds vision to its low-cost V4 Flash model
DeepSeek added multimodal input and visual-agent capabilities to V4 Flash, keeping the model in its low-cost tier while extending it beyond text.
Open-weight models lead in Aikido's latest cyber vulnerability test
In Aikido's fresh bug-finding test, DeepSeek V4 Pro recovered 28 of 32 vulnerabilities, and three cheap open-model runs together beat a single pass from Opus 5 or Grok on coverage.
Nvidia's AVO agent clears all public ARC-AGI-3 levels
Nvidia's AVO agent system completed all 183 public ARC-AGI-3 levels, illustrating how much the harness wrapped around a model can matter for long-running agentic tasks.
Replit's new Free Mode pairs a usage allowance with GPT-5.6 Luna
Replit's new Free Mode grants users a fresh usage allowance every five hours, with OpenAI's GPT-5.6 Luna handling the building work.
Slack demonstrates Slackbot as an open MCP client
Slack showed Slackbot working as an open MCP client in on-demand sessions, pulling Salesforce dashboards and a team's own internal tools directly into a channel.
Hugging Face reportedly fielding takeover offers worth $13bn or more
Hugging Face is said to be fielding acquisition interest worth at least $13 billion, nearly triple its 2023 funding-round valuation, with speculation centred on Nvidia as a natural buyer given the platform's role hosting more than two million models and 1.5 million datasets.
Nvidia's $6bn Poolside deal aims at a frontier open coding model
Nvidia agreed a $6 billion non-exclusive licensing deal plus a $1 billion equity investment in Poolside, bringing over 100 of its engineers onto Nvidia's Nemotron line after Poolside lost a 40,000-GPU cluster it couldn't finance; the stated goal is a frontier open coding model to rival DeepSeek and Moonshot.
Nvidia raises AI chip prices by as much as 17%
Nvidia has told customers that prices for its Grace Blackwell and Vera Rubin chips will rise by up to 17%, even on orders already placed for next year, with a full 72-chip Vera Rubin rack now expected to cost around $8 million; Bloomberg ties the rise to spiralling memory costs.
Alibaba raises $10bn in record Hong Kong share sale tied to its low-cost LLM push
Alibaba raised $10 billion in the largest secondary share sale of its kind on the Hong Kong exchange, though shares fell as much as 10% on the news; commentators including Michael Burry called the low-cost LLM strategy behind the raise disruptive but warned continued share issuance will keep pressuring returns.
Unitree Robotics surges 460% on Shanghai IPO debut
Unitree Robotics raised $900 million at a $9 billion market cap on the Shanghai Stock Exchange, then surged more than 460% on its first trading day, a debut Bloomberg Intelligence called a sign of strong appetite for China's embodied-AI sector.
The "AI model tier list" debate: stacks and routers now matter as much as frontier capability
A viral tier list from creator Theo (Fable 5 alone in S tier) became the occasion for a wider argument that the industry has moved past "which model is best" into a period of model stacks and routers; AT&T is shifting the share of its roughly 100,000 employees' AI queries handled by open models from 40% to a targeted 60-70%, cutting its AI coding costs 56% via a router, while Vercel gateway data shows open-weight token share jumping from 28% to 62% in two months even as OpenAI and Anthropic's combined usage kept growing.
Nvidia's first Groq 3 LPU benchmarks raise questions about its $20bn bet
The Register examined early Gemma 4 31B performance tests on Nvidia's Groq 3 LPU dataflow accelerator, describing the results as a best-case scenario for the chip that still leaves open how much the $20 billion wager on the platform will pay off.
Canonical backs project to translate legacy C into safe Rust using AI
Canonical is funding Bristol University researchers testing whether AI can reliably translate large bodies of mature C code into memory-safe Rust, an early test of AI-assisted language migration at scale.
LinkedIn users hit the "AI slop" flag button more than a million times in three weeks
Users of the world's largest professional content platform flagged AI-generated "slop" posts over a million times in three weeks, underlining how visible the backlash against low-quality AI content has become even on LinkedIn's own feed.
Original Task Manager author tests whether AI coding is ready for primetime
The original author of the 1996 Windows Task Manager utility told The Register he rebuilt it with AI tools to see whether AI coding has matured or would "degrade to slop", offering a first-hand comparison between then and now.
Microsoft moves AI governance from written policy to runtime enforcement
Microsoft has shifted its AI governance model from documented policy to controls enforced at runtime, checking agent behaviour as it executes rather than relying on review before deployment.
Anthropic's Claude Tag update lets its Slack agent join in unprompted
Anthropic's enterprise product head Scott White set out a "multiplayer AI" thesis: the updated Claude Tag reads a full Slack conversation and can insert itself without being addressed, moving away from the single-user chatbot model toward agents that hold organisational context.
IBM's next mainframe chip runs Arm and Z workloads on the same cores
IBM detailed the first dual-architecture mainframe processor, letting enterprises run Arm-native Linux software, including modern AI frameworks, alongside z/OS transaction processing on the same silicon.
JetBrains' Junie agent now runs entirely offline
JetBrains has made its Junie coding agent capable of running fully locally, with the practical caveat that it wants a 64 GB M5 Mac to do so at usable speed.
Anthropic's week-old Playground compared favourably with OpenAI's incumbent
A hands-on comparison found Anthropic's newly launched Playground outperforming OpenAI's six-year-old equivalent on developer experience, a sign of how quickly tooling around model access is being rebuilt.
Martin Fowler's Fragments notes the absence of AI whistleblowers
Fowler's 24 August Fragments covers the lack of AI whistleblowers, what could be salvaged if the AI bubble bursts, and agentic programming practice at Zalando.
"Coding expertise is going to collapse from AI reliance" tops Hacker News
An essay arguing that heavy AI reliance prevents developers building expertise reached 549 points and 538 comments, one of the week's largest discussions on what agent use does to skill formation.
Responsible AI adoption is a workflow design problem, not a policy document
Stack Overflow argues organisations cannot solve shadow AI with a document employees read once, and must instead make the sanctioned path easier than the improvised one.
"Tokenmaxxing" was always the wrong measure of AI productivity
DevOps.com traces the short life of "tokenmaxxing" and argues token consumption measures activity rather than delivered value, with security and quality costs left out of the count entirely.
LeadDev argues AI roadmaps should be built from reversible bets
Bhavuk Jain, a tech lead in applied AI at Google, argues the "right way" to build AI features shifts roughly every three months, and recommends sorting decisions by reversal cost, presenting direction as named assumptions, and shipping the thinnest end-to-end slice before committing.
General Intuition in talks at a $6bn valuation as it moves into robotics
General Intuition, building a foundation model that teaches generalised agents to move through space and time, is in talks to raise at a $6 billion pre-money valuation from Valor Ventures, Point72 Ventures and Seven Seven Six.
Nothing here matches that. Try a broader term, another day, or clear the topic filter.
The archive
Quieter
Busier