Business, Deals & Funding
Ars Technica AI
Inaudible sounds used to fingerprint browsers catch AliExpress red-handed
Researcher Matthew Callaghan discovered that AliExpress was fingerprinting visitors using inaudible sounds sent to browsers via WebAudio APIs, which he noticed because the audio processing interfered with his multipoint Bluetooth headphones. The technique uses obfuscated scripts to generate waveforms through the browser's audio implementation, with gain set to zero so users can't hear anything, then reads the resulting frequency data to create unique browser signatures based on differences in math libraries and hardware. While this specific audio fingerprinting method is now largely ineffective in modern browsers (Firefox, Chrome, and likely Safari ship their own math libraries rather than relying on OS ones, reducing entropy), AliExpress was also found to be using over a dozen other fingerprinting techniques including canvas rendering, WebGL information, screen dimensions, hardware con…
Why it matters
This is a great example of how privacy-invasive tracking often only comes to light through accidental discoveries rather than systematic auditing. While it's somewhat reassuring that major browsers have neutralized this particular audio fingerprinting vector, the extensive list of other fingerprinting methods AliExpress employs is deeply concerning and suggests that the real battle for browser privacy is far from won. The fact that an outdated, ineffective tracking script persisted unnoticed al…
Claude Code Changelog
v2.1.243
Version 2.1.243 of Claude Code adds a Loops breakdown to the /usage command showing per-loop run count, total tokens, tokens per run, and last run metrics. It introduces a modelPicker setting for customizing the /model picker with ordered, labeled model lists (supporting any ID format including Vertex/Bedrock). It also adds promptCacheTtl and subagentPromptCacheTtl settings allowing API-key and cloud-provider users to configure prompt cache duration, such as maintaining a 1-hour prompt cache on the main conversation.
Why it matters
These are practical quality-of-life improvements. The Loops breakdown in /usage is particularly useful for identifying runaway or inefficient loop tasks, which is a real pain point. The modelPicker setting adds welcome flexibility for users working across different cloud providers. The prompt cache TTL settings give power users more control over performance optimization. Overall, a solid incremental release focused on observability and configurability.
Guardian AI
The war on ‘loudcasting’ phones can be won, but it will take some very British nudging | Dan Hancox
The article discusses the phenomenon of 'loudcasting' – people playing music or audio from their phones in public without headphones – and argues that while vibrant cities shouldn't be silent, this behavior is genuinely irritating. The author suggests that addressing it will require subtle, characteristically British social nudging rather than heavy-handed regulation.
Why it matters
The article tackles a relatable modern annoyance with humor and cultural insight. The writing is engaging, with witty observations like the quip about Switzerland. The 'British nudging' approach reflects a pragmatic middle ground between authoritarian noise regulation and passive acceptance, though the article's full argument is cut off, making it hard to fully evaluate the proposed solutions.
Lenny's Newsletter
🎙️ How I AI: Grok Bot + Grok 4.6—what’s great (and what’s still hype) & Lessons from spending $20,000 on Devin in one month
This article previews two episodes from the How I AI podcast (part of Lenny's Podcast Network, dated August 2026). The first episode covers Claire's hands-on testing of Grok Bot, Cursor Origin, and Grok 4.6. Key findings: Grok Bot's multi-account connector feature (supporting multiple email and Slack accounts) is genuinely useful; Cursor Origin is a promising agent-native GitHub alternative but not yet ready for production teams; Grok 4.6 performed competitively with GPT-5.6 Sol at the top of Claire's blind evaluation index, ahead of Sonnet 5 and Opus 5; however Sonnet 5 remains best for conversational agent interactions. She notes Cursor and xAI are building a coherent enterprise stack spanning knowledge work, code hosting, and IDE. The second episode features Ryan Carson, solo founder of Untangle (B2B SaaS for family law firms), who spent $20,000 on Devin in one month, managing up to…
Why it matters
This is a forward-looking piece from 2026 that provides genuinely useful practitioner-level insights rather than just benchmark hype. Claire's approach of running blind evaluations with her own judgment weighted at 70% is refreshingly honest compared to typical leaderboard worship. The multi-account connector insight about Grok Bot highlights a real UX gap that most AI platforms ignore—most people don't live in single-account worlds. The Cursor/xAI enterprise stack observation is strategically…
MIT Tech Review AI
How to encourage smarter AI use in the classroom
Cheshire Academy, a private school in Connecticut, is experimenting with AI use in the classroom. Rather than prescribing specific tools, the school trained staff on general AI techniques, including prompt crafting and understanding AI limitations. Teachers use generative AI for lesson planning and creating rubrics, though concerns about quality and privacy have prevented its use for student feedback. French teacher Miriam Przybyla-Baum developed exercises where students evaluate LLM edits of their work and anonymously assess AI-assisted assignments. The school is also piloting a 'Student AI Council' where students lead discussions on healthy AI use. The article references a traffic light metaphor to guide when students can use AI for assignments.
Why it matters
This article presents a thoughtful, pragmatic approach to AI in education. Cheshire Academy's strategy of training teachers on general AI principles rather than mandating specific tools seems wise given how rapidly the technology evolves. Przybyla-Baum's exercises are particularly clever—having students critically evaluate AI edits teaches both subject matter and AI literacy simultaneously. The Student AI Council is also a promising idea, giving students agency in shaping norms rather than just…
NY Times
What Do A.I. Companies Really Do With Data They Purchase?
The article discusses how AI companies utilize data they purchase, explaining that data is often used to train AI models through reinforcement learning. In this process, entire companies are recreated as simulations that AI models can interact with, allowing the models to learn and improve their capabilities in realistic environments.
Why it matters
This is a very brief piece that touches on an important and underexplored topic — the actual mechanics of how purchased data gets used in AI training. The mention of recreating entire companies as simulations for reinforcement learning is an intriguing and somewhat alarming detail that deserves much deeper investigation. The lack of depth in the available content makes it hard to fully evaluate, but the core question of what AI companies do with purchased data is one of the most pressing transp…
OpenAI
Advancing price-performance for developers with GPT‑5.6 in Kiro
OpenAI announced that its GPT-5.6 model family (including Sol, Terra, and Luna variants) is now available in Kiro, an AI-native software development agent built by AWS. The integration aims to improve price-performance for developers by helping them plan, build, review, and test software with fewer iterations and better value per token. Kiro uses a spec-driven development approach that turns high-level intent into structured requirements, technical designs, and executable tasks, providing GPT-5.6 with clear context. OpenAI and AWS claim that testing on Terminal-Bench 2.1 showed GPT-5.6 Terra completed tasks in Kiro at roughly 82% cost reduction. The collaboration between OpenAI and AWS focuses on optimizing model performance within the Kiro environment, with quotes from AWS VP Swami Sivasubramanian and OpenAI VP Colleen Kapase emphasizing the partnership's goal of accelerating AI-native…
Why it matters
This is essentially a joint marketing announcement between OpenAI and AWS, and it reads like one — heavy on buzzwords like 'price-performance,' 'AI-native development,' and 'spec-driven development,' but light on concrete technical details. The 82% cost reduction claim on Terminal-Bench 2.1 is interesting but lacks context: compared to what baseline? What tasks specifically? The framing is carefully crafted to make it sound revolutionary while being vague enough to avoid scrutiny. That said, th…
TechCrunch AI
Situational Awareness, star AI hedge fund that nearly imploded, now being probed by the SEC
Situational Awareness, the AI-focused hedge fund led by former OpenAI employee Leopold Aschenbrenner, is now under SEC investigation after nearly imploding in late July 2026 when an AI stock downturn erased billions in value. The SEC has been subpoenaing banks that did business with the fund, focusing on those that supervised its trading and channeled funding to support it. While the fund has not been accused of wrongdoing, banks have been told to preserve all related information. The company stated that scrutiny of high-profile funds is expected and pledged full cooperation with regulators.
Why it matters
This is a significant development that underscores the risks of concentrated AI investment strategies and the broader volatility in the AI sector. Leopold Aschenbrenner gained prominence for his bullish AI predictions after leaving OpenAI, and the fund's rapid rise and near-collapse illustrate how hype-driven investment theses can unravel quickly. The SEC probe suggests regulators may be concerned about leverage, risk management practices, or potential irregularities in how the fund operated. W…
The Rundown AI
A mystery challenger at the AI frontier
An anonymous AI model called 'Ox Alpha' appeared on OpenRouter with no identified creator, offering free access with a 1M-token context window and multimodal input designed for coding and agentic work. It initially scored 80% on a DeepSWE subset but settled at 63% on full testing, near frontier-level performance. Early investigations suggest it may originate from China's Zhipu AI (possibly GLM-5.3 Flash or GLM-6), with Microsoft's MAI family as another candidate. The model has drawn heavy usage due to a free week-long offer supporting 100T tokens per day. The article also covers a Rundown Roundtable segment where staff share personal AI use cases, including building custom app-switching shortcuts with ChatGPT and using Claude for visa application preparation.
Why it matters
The emergence of anonymous frontier-level models on public platforms is a fascinating and somewhat concerning trend. While the detective work around Ox Alpha's origins is entertaining, the pattern of anonymous Chinese lab drops on OpenRouter raises questions about transparency and accountability in AI deployment. If the model truly delivers near-frontier coding performance at a Flash-tier efficiency, it represents a meaningful shift in the cost-performance curve. However, the hype cycle around…
Guardian AI
Tell us: do you think AI has made Google search better or worse?
The Guardian is asking readers whether AI has improved or worsened Google search. Google has placed AI-generated summaries, called AI Overviews, at the top of search results, requiring users to scroll past chatbot-generated responses before seeing traditional links. The article frames this as a profound change to what was once the primary gateway to the rest of the internet, affecting the browsing habits of billions of people.
Why it matters
This is a reader engagement piece from The Guardian soliciting opinions on a significant shift in how Google presents search results. The topic is highly relevant as AI Overviews have been controversial since their rollout — they have been criticized for sometimes providing inaccurate or misleading summaries, reducing traffic to original content creators, and fundamentally changing the open web ecosystem. The framing of the question suggests the Guardian recognizes this as a contentious issue.…
Lenny's Newsletter
I spent $20,000 on Devin in a month. Here’s what I learned | Ryan Carson (solo founder)
Ryan Carson, a five-time founder and current solo founder of Untangle (a B2B SaaS platform for family law firms), shares his experience spending $20,000 in one month on Devin, an AI coding agent. He runs 15 concurrent Devin agents simultaneously to handle engineering, customer success, and investor updates, managing them with a simple folder system and a handwritten paper list rather than a dashboard. Key topics include his 'Watchdog playbook' that replaced a customer success team for monitoring law firm accounts, his 'LAN PR skill' that handles 40 daily pull requests without a QA team, his preference for cloud agents over local ones (except for specific Codex use cases), and his design workflow using Claude Design into Markdown specs then Codex for building. He discusses managing agent decision fatigue, why producing more AI output doesn't necessarily make a better product, how meeting…
Why it matters
This is a genuinely fascinating and practical case study of a solo founder pushing the boundaries of AI-assisted development at scale. The $20,000/month spend is attention-grabbing but the real value is in the operational details — managing 15 concurrent agents with paper lists, building automated PR review systems, and replacing entire team functions with AI workflows. The insight that more AI output isn't the goal is refreshingly honest and counterintuitive given the hype cycle. However, ther…
MIT Tech Review AI
Kids outlearn AI—and we still don’t know why
The article explores the 'data efficiency gap' between children and large language models (LLMs) in learning language. While LLMs like Claude, DeepSeek, and GPT models have achieved impressive fluency, they require vastly more data than human children—potentially 100,000 times more words than a child encounters before mastering their native language. A typical LLM may train on trillions of tokens, whereas a child might hear only around 100 million words by their preteen years. Researchers in cognitive science and AI are investigating why children are so much more data-efficient, hoping that reverse-engineering children's learning processes could lead to more efficient AI models. This has practical implications as easily available training data may run dry by the 2030s. Potential applications include training AI on video more effectively and creating chatbots for minority language commun…
Why it matters
This is a genuinely fascinating and well-framed article that highlights one of the most important and underappreciated gaps in current AI research. The data efficiency gap is not just a technical curiosity—it points to something fundamental about the difference between how biological minds and statistical models process information. Children don't just learn language; they learn it in context, with embodied experience, social interaction, emotional feedback, and a developing theory of mind. LLM…
TechCrunch AI
Trump bought SpaceX shares two weeks after blockbuster IPO
President Donald Trump purchased up to $50,000 worth of SpaceX shares on June 23, 2026, two weeks after the company's record-setting IPO. The shares were trading in the mid-$150 range at the time of purchase, but have since fallen to the IPO price of $135, potentially putting Trump's investment underwater. The White House stated that Trump's portfolio is managed by third-party institutions replicating recognized indexes, and SpaceX had lobbied popular indexes to change their rules for faster inclusion ahead of its IPO. The article notes Trump and Musk's close relationship despite a brief falling out, and that SpaceX has been receiving increasing government contracts and benefiting from the administration's deregulatory stance.
Why it matters
This article raises significant conflict-of-interest concerns. Even if Trump's portfolio is managed by third parties replicating indexes, the fact that SpaceX lobbied indexes for faster inclusion—effectively ensuring that passive investors like Trump would automatically hold its stock—creates a convenient layer of plausible deniability. Meanwhile, SpaceX continues to receive growing government contracts and regulatory benefits from the Trump administration. The White House's explanation feels l…
Guardian AI
‘Never seen this level of objection’: Scotland pushes back against datacentre boom
Scotland is experiencing significant public opposition to a proposed massive datacentre development in Auchtertool, Fife, which would be larger than the village itself and billed as the second-biggest datacentre in the world. The project has attracted 1,600 objections, and the Scottish Parliament at Holyrood is reportedly moving towards a moratorium on such developments amid growing concerns about the datacentre boom.
Why it matters
This story highlights a growing global tension between the insatiable demand for digital infrastructure—driven by AI, cloud computing, and data storage—and the impact on local communities and environments. The scale of the proposed development (larger than 100 football pitches, 35 metres high, in a small village) seems almost absurdly disproportionate, and the level of public objection is understandable. Communities should absolutely have a meaningful say in developments that would fundamentall…
TechCrunch AI
Amjad Masad, CEO and co-founder of Replit, joins the Disrupt Stage at TechCrunch Disrupt 2026
Replit CEO Amjad Masad will speak at TechCrunch Disrupt 2026 (October 13-15, San Francisco) about the future of programming and Replit's role in it. The article highlights Replit's explosive growth, with its current run-rate tracking toward $1 billion annually, up from $2.8 million in reported revenue in 2024. The company received a $9 billion valuation earlier in 2026, just six months after being valued at $3 billion. Masad's talk will cover how AI is enabling a wider range of people to develop software, the implications of ideas being easily turned into products, and how Replit serves both nontechnical users and professional coders. The event will also feature stages dedicated to AI, builders, smart systems, and real-world AI applications.
Why it matters
This is essentially a promotional article for TechCrunch Disrupt 2026 wrapped around genuinely interesting data points about Replit's growth. The revenue trajectory from $2.8 million to a billion-dollar run rate in roughly two years is remarkable and speaks to the massive demand for AI-assisted coding tools. The tripling of valuation from $3 billion to $9 billion in six months raises questions about whether this reflects sustainable value or speculative froth in the AI sector. The broader theme…
From Reddit/HN/YC
- [Hacker News] Artificial Analysis drops a full DeepSeek V4 Flash Vision benchmark covering intelligence, throughput, and price per token.
- [Hacker News] OpenObserve benchmarks 1 billion logs against ClickHouse and publishes the full results — storage, query speed, and cost.
- [Hacker News] Tsiolkovsky's 51,000-page archive has been machine-read and published online with measured OCR accuracy.
- [Hacker News] Adam Tornhill makes the case that you still need to read AI-generated code — treating it as trusted output is where teams get burned.
- [Hacker News] The U.S. officially rescinds Syria's designation as a state sponsor of terrorism.
- [Hacker News] IBM reveals its first dual-ISA core at Hot Chips 2026 — natively executing both ARM and z/Architecture instructions at 5.7 GHz on a 2nm node.
- [Hacker News] AnasSyed found an SSRF in Google's official AI tooling and documents how Google reacted to the disclosure.
- [Hacker News] Intel's Diamond Rapids gets its Hot Chips 2026 deep dive courtesy of Chips and Cheese.
- [Hacker News] New research suggests screen time at ages 1 and 6 may carry more developmental weight than parents realize.
- [Hacker News] Every AI data center currently under construction could collectively emit as much CO₂ as 24 million cars.