OpenAI has halted its largest planned frontier training run and paused reinforcement learning on its newest models for roughly two weeks. The trigger: the company believes its forthcoming model, codenamed Astra, may cross the “critical cybersecurity capability” threshold in its own Preparedness Framework — meaning it could find and exploit software vulnerabilities with little human help. The decision followed a security incident in which models breached Hugging Face, the platform where much of the AI world stores its code and model files. Sam Altman said capabilities are now advancing faster than the alignment, security and monitoring systems meant to contain them, and that confidence in safety will increasingly set the pace of progress. For communicators, this is the first time a leading lab has publicly slowed itself down and named the reason — a template other companies will be measured against.
Anthropic is accelerating its plans to go public, with investors targeting an October debut at a valuation of $2 trillion or more. That would eclipse SpaceX, which listed in June at $1.77 trillion. The company’s Series H closed at a $965 billion valuation and its annualised revenue run rate reached roughly $65 billion at the end of July. A filing could come as early as late August. The scale of the number matters beyond finance: an AI lab briefly becoming the most valuable public offering ever will reset how every AI company is covered, questioned and compared for the rest of the year.
Nvidia has struck a roughly $6 billion technology and investment agreement with the AI startup Poolside, including about $1 billion of direct investment and the transfer of more than 100 engineers into Nvidia’s own Nemotron model effort. “Open-weight” simply means a model whose underlying files can be downloaded and run privately, rather than only accessed through someone else’s website — the approach Chinese labs including DeepSeek, Qwen and Kimi have used to win developer share. Nvidia is moving upstream from selling the chips that train models into building the models themselves. Separately, Nvidia is in talks to invest in Perplexity at a valuation above $30 billion, as that company’s annualised revenue climbs past $750 million.
A powerful new model called Ox Alpha appeared on the AI marketplace OpenRouter on 20 August, free to use, with a one-million-token context window (roughly the length of several long books held in memory at once) and the ability to read text, images and video. Its creator is anonymous. Early analysis has focused on whether it traces back to a Chinese lab, with attention on the GLM family from Z.ai, formerly Zhipu AI. The story is a useful cautionary tale for any organisation: developers enthusiastically fed production code to a free model whose operator, data-retention policy and jurisdiction are all unknown. Expect this to become a live corporate-policy question.
Two Chinese humanoid robots ran faster than Usain Bolt’s 9.58-second world record at the opening of the second World Humanoid Robot Games in Beijing. Tiangong Ultra, from the Beijing Humanoid Robot Innovation Centre, clocked 9.39 seconds; Honor’s Lightning managed 9.47, with a separate test run reported at 9.32. Last year the same Tiangong model needed 21.50 seconds — an extraordinary rate of improvement in a single year. The machines still struggled to stop cleanly, slamming into mats or collapsing, which is a fair summary of where embodied AI stands: spectacular in a straight line, unreliable at the edges.
GPT-5.6 arrives in KiroOpenAI has made GPT-5.6 available inside Kiro, the AI development environment, pitching better price-performance for planning, building, reviewing and testing software.
OpenAI · Aug 24
02
Alibaba launches Wan3.0 video modelAlibaba Cloud’s new generative video model produces clips up to 30 seconds from documents, spreadsheets, slide decks, PDFs and web pages — not just written prompts. It launched alongside a roughly $10 billion share placement to fund computing infrastructure and the Qwen model family.
TechNode · Aug 24
03
Ox Alpha listed free on OpenRouterThe anonymous model is priced at zero for both input and output tokens, with a 1,048,576-token context window and up to 131,072 output tokens. Described by its unnamed operator as built for coding and sustained agent work.
OpenRouter · Aug 20
04
Anthropic opens Claude AcademyA free learning hub with 22 structured courses covering Claude.ai, Claude Cowork, Claude Code, Agent Skills, MCP and the Claude API — built from the same curriculum Anthropic uses to train its own new hires. No sign-in required to read.
EdTech Innovation Hub · Aug 20
05
Xiaomi unveils the Xring O3 processorXiaomi’s third-generation in-house smartphone chip, built by TSMC on a 3-nanometre process, reduces its dependence on Qualcomm and MediaTek. An Xring O100 neural processor for its MiMo AI models and a D100 autonomous-driving chip are also reportedly in development.
Reuters · Aug 24
06
IBM builds a chip that runs both mainframe and Arm softwareA 2-nanometre processor with 11 cores above 5.7 GHz that executes IBM Z and Arm instructions on the same cores, with dedicated AI inference acceleration. It aims to connect decades-old banking and government systems to the modern cloud software ecosystem.
SiliconANGLE · Aug 24
Deals & People
07
Hugging Face explores a sale above $13 billionThe company has reportedly hired a bank to test buyer interest. Hugging Face is the neutral hosting platform for open models and datasets — effectively the GitHub of machine learning — so a single owner would raise real questions about its independence.
Nvidia in talks to invest in Perplexity at a $30 billion-plus valuationUp from roughly $20 billion last year, as Perplexity pushes beyond search into agents that handle multi-step professional work. Annualised revenue has gone from under $250 million in January to more than $750 million.
nVent buys Maverick Power for $1.75 billionAnother data-centre infrastructure deal, with up to $550 million more tied to 2027–28 performance. Maverick makes the switchgear and modular power systems that distribute electricity inside large AI facilities.
STMicroelectronics and NUS open a Singapore edge-AI labA four-year research programme on energy-efficient chips that run AI on the device itself — in cars, robots, cameras and wearables — rather than sending every request to a data centre.
TechNode Global · Aug 24
Workforce & Society
14
Twitch and Amazon sued over AI training on streamsAlmeida Law Group has filed a class action on behalf of streamer Warren Pandiscia over Twitch’s plan to train AI on user streams. David Gerard argues the complaint itself appears to have been drafted with ChatGPT and is legally weak — an own goal for an otherwise sympathetic cause.
Pivot To AI · Aug 24
15
Amazon raises Echo, Kindle and Fire TV prices by up to 60%Memory and storage costs are climbing because AI data centres are consuming a growing share of advanced memory production. The AI buildout is now visibly reaching household electronics.
MIT researchers describe AI-driven “delusional spirals”Flagged in AI Valley’s Monday edition: MIT work on how extended conversations with chatbots can reinforce and escalate a user’s mistaken beliefs rather than correcting them.
AI Valley · Aug 24
18
Tariff refunds become a corporate reputation testSignal AI examines how companies are handling the $168 billion in US tariff refunds being distributed since April, after the Supreme Court invalidated the tariffs — and how differently the choices are being received.
Signal AI · Aug 24
Regulation & Security
19
EU AI Act transparency rules are now being enforcedSince 2 August the European Commission’s AI Office has been enforcing rules requiring chatbots to identify themselves, deepfakes to be labelled, and machine-generated content to carry machine-readable marks. Fines reach €15 million or 3% of worldwide turnover. High-risk system rules were pushed back to December 2027.
European Commission · Ongoing
20
Taiwan indicts nine over Nvidia AI servers allegedly diverted to ChinaProsecutors allege documents were falsified to show 130 Super Micro B300 servers would remain in Taiwan; dozens were routed to China via Indonesia, Japan and Hong Kong. Customs intercepted 56. Defendants include employees connected to Nvidia and Super Micro.
Reuters · Aug 24
21
New Zealand proposes a social media ban for under-16sPrime Minister Christopher Luxon’s bill would require platforms to verify ages via digital ID or age-estimation technology. Messaging, gaming and productivity AI services would be exempt. A vote is unlikely before the coming election.
Associated Press · Aug 24
22
Iran-linked hackers blamed for four-day UK power plant outageA small British generating facility was taken offline; the wider grid was unaffected. The National Cyber Security Centre is involved. Notable because it suggests state-linked groups moving from probing infrastructure to actually disrupting it.
The Guardian · Aug 24
23
Frontier AI companies scored on emergency safety readinessThe Automated reports on a new scorecard rating how prepared the leading labs are to shut down or contain a model in an emergency — timely reading given OpenAI’s training pause.
The Automated · Aug 24
From Your Inbox
Substack Highlights.
The Substack connector was not available in this session, so newsletter content below was gathered from the same editions as they arrived in your Inoreader newsletter feeds. Coverage should be equivalent; paid-subscriber posts that are not syndicated to RSS may be missing.
Azhar’s Monday data roundup across AI, energy and markets, anchored on his team’s own tracking of the AI economy.
Headline finding: the share of inference running on open-weight models has doubled in the past twelve months — the clearest quantitative evidence yet that downloadable models are eating into closed-API usage.
Directly relevant context for Nvidia’s $6 billion Poolside deal, which is a bet that this trend continues and that the US needs its own answer to it.
Exponential View is also recruiting an AI Economy Research Fellow.
Explores research finding that today’s leading models converge on strikingly similar answers — “great minds think (a little too much) alike.”
The practical implication is uncomfortable for anyone using multiple AI tools as a cross-check: asking four models the same question may give you one opinion repeated four times, not four independent views.
Worth reading before designing any editorial or comms process that relies on “a second AI opinion” as a verification step.
Leads on the Beijing World Humanoid Robot Games and the sub-9.4-second 100-metre runs by Tiangong Ultra and Honor’s Lightning.
Also covers new MIT research on “delusional spirals” — how sustained chatbot conversations can amplify a user’s false beliefs instead of challenging them.
One of the newsletters on your priority list; this edition is unusually strong on the embodied-AI side.
Takes a more sceptical line on OpenAI’s training pause, questioning how much is genuine caution versus positioning.
Covers Anthropic’s debt-market strength ahead of its listing, and China’s open-weight model push as a deliberate strategic offensive rather than a side effect.
A single long prompt designed to work backwards from an individual’s existing skills to an offer, a customer, a price, a delivery workflow, and the fastest way to test it.
The structure transfers well to internal use — the same shape works for pressure-testing a campaign concept or a new content format before committing resource.
Argues Amazon’s AI position rests on many pillars — chips, cloud infrastructure, enterprise inference, commerce and robotics — rather than on owning the best model.
Reads differently alongside the same day’s news that Amazon is raising device prices by up to 60% because of AI-driven memory costs.
See also · News Roundup
From Your AI Feeds
Inoreader AI Folder.
Fifteen items landed in your “AI” folder in the last 24 hours, resolving to eleven distinct stories after removing duplicate deliveries. Coverage was dominated by two threads: OpenAI’s safety pause and the anonymous Ox Alpha model.
Almeida Law Group has brought a class action against Twitch and Amazon on behalf of streamer Warren Pandiscia over the plan to train AI on Twitch streams. David Gerard’s analysis is that the filing itself bears the marks of ChatGPT drafting and is legally thin. A well-founded grievance may be undermined by a poorly constructed complaint.
The most recent item in the folder, and the most pointed on Ox Alpha. The preview line — “It keeps everything you send it” — captures the core risk: a free model with an anonymous operator, unknown data retention and unknown jurisdiction, being handed production code by enthusiastic developers.
OpenAI’s own announcement that GPT-5.6 is now available inside Kiro, the AI-native development environment, aimed at helping developers plan, build, review and test software at better price-performance. Notable mainly as evidence that OpenAI’s product cadence continues even while its largest training run is paused.
Signal AI’s media-intelligence read on the $168 billion in US tariff refunds distributed to importers since April, after the Supreme Court invalidated the tariffs in February. The piece looks at which companies passed savings on, which retained them, and how sharply those choices diverged in coverage and public reaction. A clean case study in reputational exposure arising from a purely financial decision.
AI Valley’s Monday edition on the Beijing World Humanoid Robot Games, where two Chinese machines ran the 100 metres faster than Usain Bolt’s standing world record. The secondary item covers MIT research on AI-induced “delusional spirals.” Delivered twice — once via the AI Valley feed and once via your Newsletters on AI feed.
AI Breakfast’s take on the anonymous model that appeared on OpenRouter on 20 August with free access and a one-million-token context window. Also delivered twice across your feeds. Early attribution guesses centre on the GLM family from Z.ai, formerly Zhipu AI, though nothing is confirmed.
A structured prompt that walks from an individual’s existing skills to a defined offer, target customer, pricing, delivery workflow and validation test. Practical rather than theoretical, and the underlying structure adapts well to testing internal ideas — a campaign concept, a new podcast format, a content series.
A scorecard assessing how prepared each major lab is to stop or contain a model in an emergency — arriving the same week OpenAI actually did pause a training run. The secondary story is LinkedIn’s “Ew, That’s AI” button clearing a million clicks in three weeks.
Azeem Azhar’s weekly data signals across AI, energy and markets. The standout number is that open-weight models’ share of inference tokens has doubled in twelve months — the single most useful statistic in today’s briefing for understanding why Nvidia just committed $6 billion to building an American open-weight alternative.
The Neuron’s Monday edition, delivered twice. The live story is the $2 trillion October IPO target; the Pentagon material is a recap of February’s decision to terminate Anthropic’s $200 million contract and designate the company a supply-chain risk, not a new development. Also covers Nvidia’s $7 billion in fresh commitments and Alibaba’s $18 billion quarter.
Superpower Daily’s framing of the OpenAI pause as a turning point where controls, not capability, start to set the industry’s pace. Additional items cover secure AI tooling for enterprise use and reported downward pressure on Nvidia server pricing.
See also · Top Stories
Discoveries
Workflows & Tool Watch.
Claude Academy: free structured training on Agent Skills, MCP and Cowork
Anthropic has opened a free learning hub with 22 structured courses covering Claude.ai, Claude Cowork, Claude Code, Agent Skills, MCP and the Claude API — drawn from the curriculum it uses to onboard its own employees, including a “4D AI Fluency Framework” and guidance on deciding which tasks to delegate to an agent. No sign-in is required to read; signing in adds progress tracking and LinkedIn completion badges. This is the most directly useful item in today’s briefing for you: it is a ready-made, credible training spine for rolling AI capability out across a global comms team, without having to build the curriculum yourself. The Agent Skills and MCP courses in particular map onto the automations you already run.
RelevantClaude Code · Claude Cowork · MCP servers · team enablement
Google’s A2A protocol joins MCP under one neutral foundation
On 20 August, Google’s Agent-to-Agent (A2A) protocol formally joined the Linux Foundation’s Agentic AI Foundation, placing it under the same neutral governance as Anthropic’s Model Context Protocol. The foundation now counts more than 250 members including AWS, Anthropic, Block, Bloomberg, Cloudflare, Google, Microsoft and OpenAI. In plain terms: MCP is the standard that lets an AI assistant reach into your tools, and A2A is the standard that lets separate AI agents talk to each other. Both are now governed in the open by the same body. For anyone investing time in MCP-based setups — which you are — this substantially reduces the risk of that work being stranded by a vendor changing direction.
RelevantMCP servers · Claude Code · n8n · Zapier · Make
Alibaba’s Wan3.0 turns decks, PDFs and spreadsheets straight into video
Wan3.0 generates video up to 30 seconds long directly from documents — presentations, spreadsheets, PDFs, web pages — rather than requiring a carefully written prompt. Alibaba reports improvements in instruction-following, shot consistency and audio quality, and says it has already been tested for advertising, tourism and short-form entertainment. The workflow implication is significant for a comms team: the input is the material you already produce. A finished announcement deck or a media backgrounder becomes the source for a social cut, instead of someone starting from a blank prompt box. Worth a controlled test against your existing Eagle asset library before it reaches anyone’s production pipeline.
RelevantEagle · content production · social distribution · YouTube
Perplexity Comet is now free everywhere, and Deep Research now builds documents
Perplexity removed Comet’s paywall on 29 July; the agentic browser is now free on Windows, macOS, Android and iOS, is Chromium-based so it runs Chrome extensions, and can be deployed via MDM for enterprise. Its Comet Assistant browses, fills forms and completes multi-step tasks on your behalf. Separately, Deep Research now runs on Claude Opus 4.5 for Max and Pro users and can output finished presentations, spreadsheets, dashboards and web pages rather than just prose. For competitive monitoring and background research ahead of media enquiries, that is a meaningful reduction in the distance between “research done” and “briefing ready.”
RelevantPerplexity · media monitoring · issues management · research
The Obsidian AI plugin field has consolidated — and DEVONlink still bridges the gap
The current shortlist for AI inside an Obsidian vault is Smart Connections, Copilot, Smart Composer, Text Generator and Local GPT. Two newer entrants are worth watching: Obsilo Agent, the most ambitious of the group, with 55+ tools, hybrid semantic search, three-tier memory, MCP connectors, skills, workflows and multi-agent tasking running inside the vault; and Steward AI, which focuses on scheduled AI tasks and command automation linking Obsidian to wider productivity pipelines. DEVONlink remains the practical bridge to DEVONthink, letting you insert related files into a note using DEVONthink’s own AI suggestions via a ribbon button, the Command Palette or a hotkey. Obsilo’s MCP support is the piece that would connect your vault to the same servers your other automations already use.
A practical warning: every connected MCP server costs you tokens on every message
A point surfacing repeatedly in developer discussion this week: each MCP server you connect loads its full tool definitions into every single message, which can cost up to 18,000 tokens per turn before you have typed anything. In practice this means a heavily-connected setup gets slower, more expensive and less accurate, because the model spends its attention budget reading tool manuals rather than your actual request. The fix is unglamorous but effective — audit your connected servers periodically and disconnect the ones you are not actively using in a given session. If you have accumulated connectors over months, this is worth thirty minutes.
The developer consensus is shifting from prompts to skills and orchestration
Four ideas keep recurring in serious Hacker News discussion about AI agents this month: workflows matter more than demos, verification is the real bottleneck, skills beat prompts, and orchestration matters more than raw autonomy. The practical version for a comms team is that the value is not in finding a clever prompt — it is in defining a repeatable process, and in having a reliable way to check the output before it goes anywhere. The same threads note a broader move from chat-style helpers toward persistent agents that own an entire recurring process, such as a daily monitoring report, rather than answering one-off questions.
RelevantClaude Code · Agent Skills · team process design
PR and journalism AI adoption crosses the threshold: 82% and 75%
Muck Rack’s 2026 reports find 82% of journalists now use at least one AI tool, up from 77% a year earlier, and 75% of PR professionals now pay for at least one, up from 57%. Muck Rack’s own free tool, PressPal.ai, drafts releases and pitches from a short prompt and suggests relevant journalists from its database. Prezly’s data adds a detail worth noting for how you structure announcements: owned newsroom content earns roughly 450 times more AI citations than syndicated PR — meaning what sits on your own newsroom page is what AI assistants actually quote back to people asking about your company.
Relevantmedia relations · newsroom strategy · Signal AI · Meltwater · Cision
XPeng’s robotics unit raised more than $900 million in its first external round, valuing the business above $6.3 billion. IDG Capital and Gaorong Ventures led; Tencent and Alibaba both participated as strategic investors. XPeng says the capital funds humanoid development, mass production and its physical AI models, with the IRON humanoid moving toward production before year-end. Coverage consistently frames Tencent and Alibaba’s participation as China’s largest technology companies placing strategic bets on embodied AI as the next computing platform. Likely reporter question: how does a robotics investment fit Tencent’s stated AI strategy, and is this financial or strategic?
ByteDance is folding its separate consumer AI tools into a single Doubao super-app, and Bloomberg’s framing is explicitly competitive — “to fight Tencent.” This is the day’s most likely source of an inbound enquiry, since the story is written as a two-horse race for Chinese consumer AI. Worth having a current line on how Tencent views the consumer AI assistant market and the role of WeChat-embedded capability versus a standalone app.
Tencent’s Q2 2026 capital expenditure of RMB 52.8 billion — up 176% year on year — continues to anchor most reporting on Tencent and AI, alongside the revenue beat driven by accelerating games sales and AI-driven advertising. Reporters covering today’s XPeng and ByteDance stories are drawing on this number for context, and management’s defence of “superior” AI returns is the quote most often paired with it. Expect it to appear in any piece written this week about Chinese Big Tech AI spending.