tokenando

OpenAI published a case study and executive interview with Sam Nickless, partner and chief executive of Gilbert + Tobin, one of Australia's leading corporate law firms, on November 12, 2026, detailing the firm's multi-phase AI integration using ChatGPT Enterprise, ChatGPT Work, and Codex. The firm achieved 87% active usage among enabled ChatGPT seats as of June 2026, more than twice its typical adoption rate for new tools, without mandating usage targets. Specific workflow outcomes include a recruitment research task reduced from approximately 4 hours to 20 minutes; selected conflict, anti-money laundering, politically exposed person, and know-your-customer checks reduced from 3 to 8 hours to approximately 5 minutes; and audit report preparation across 300 entities that previously required a full day of manual work completed by Codex. A partner annual review process covering 120 partners, which would previously have required several people over multiple days, was completed using ChatGPT Work at a cost of approximately $62. The firm also uses Harvey for legal-specific workflows and has built a custom GPT functioning as a digital twin of the chief executive for internal use by direct reports. Nickless described the firm's AI journey in 3 phases: experimentation with no client data, personal productivity through ChatGPT Enterprise with Australian data residency, and organisational productivity through agentic workflows with Codex. He disclosed that the firm anticipates changes to how it charges clients and how it contracts new clients as AI compresses the time required for work previously billed by the hour. 🤖 made with AI Full details via OpenAI / openai.com/index/gilbert-tob…

Mistral chief executive Arthur Mensch publicly called on French lawmakers to reject the Darcos bill between November 8 and 10, 2026, ahead of a scheduled National Assembly debate on November 19. The bill, a single-article proposal adopted unanimously by the French Senate on April 8, 2026, would establish a civil presumption that an AI provider used a protected work whenever an indication linked to the model's development, deployment, or output makes such use plausible. Mensch's objection centres on the reversal of the burden of proof: proving that a specific work was not used in a training dataset of billions of words would, he argues, require AI companies to expose trade secrets and generate extensive litigation. He linked the legal risk to Mistral's infrastructure plans, stating that the bill would influence decisions about whether to build a planned data centre in the Essonne department south of Paris or elsewhere, and where to conduct research. As an alternative, Mensch proposed a revenue-based contribution from all AI providers operating commercially in Europe, including foreign companies, to fund a European creative sector fund. Senator Laure Darcos, who drafted the bill, replied publicly that data centres are not targeted by the presumption and that business secrecy cannot excuse a breach of property rights. A coalition of French creative organisations, including writers, journalists, musicians, and artists, published a collective tribune in Le Monde on November 10 opposing Mistral's position. The bill was blocked from reaching the full chamber on June 11 after 110 amendments made debate impossible; the November 19 session is its next scheduled opportunity. 🤖 made with AI Full details via Sifted / sifted.eu/articles/mistral-c…

Booz Allen Hamilton, a US government and defence technology contractor with a market capitalisation of approximately $10 billion, switched its cybersecurity and defence software workloads from Anthropic's models to Nvidia's Nemotron open-source models several months before November 2026. The move was reported by The Information, citing the company's chief technology officer, Bill Vass, who said the switch produced significant cost savings, which the company declined to quantify. Booz Allen spends tens of millions of dollars annually on AI. Developers at the firm continue to use Anthropic's Claude Code for software development unless customers request otherwise, and Vass indicated interest in extending Nemotron use to coding as the models improve. The switch was facilitated in part by Palantir, which launched a product in late June 2026 to help US government agencies and enterprises run Nvidia open-source models through its software. Nvidia has also signed a deal to use Palantir's software internally for supply-chain management. Separately, Nemotron accounted for approximately 17% of the roughly 50 trillion tokens processed through the open-source coding platform Kilo Code in October 2026, up from 8% the previous month. On OpenRouter, which processes approximately 700 trillion tokens monthly, Nemotron represented only about 4% of tokens. Anthropic accounted for 31% of OpenRouter customer spending in November 2026, with Nvidia below 1%. 🤖 made with AI Full details via The Information via Benzinga / benzinga.com/content/6312366…

OpenAI launched "Businesses in ChatGPT" in beta on November 10, 2026, introducing 2 linked features: Business Hub, which allows companies to claim and verify their presence in ChatGPT and access aggregated engagement analytics; and Business Agents, which allow eligible US e-commerce and online service businesses to build and publish AI-powered agents that ChatGPT users can interact with directly when a business appears in a model response. The platform reaches 1.2 billion users per week, according to OpenAI. Verification is handled through Business Hub via phone, email, website domain, or Shopify account, with a separate identity check through Persona required before publishing changes or creating an agent. Shopify merchants can connect their stores directly. Business Hub analytics cover mentions, website clicks, and discovery topics, giving marketing teams data on how AI-mediated discovery affects their customer acquisition funnel. Businesses can sponsor their Business Agent and use it as the destination for ChatGPT Ads, with sponsored placements clearly labelled. OpenAI states it does not share individual user conversations with businesses; analytics are aggregated and available only above minimum interaction thresholds. The feature is initially limited to eligible US e-commerce and online service businesses, with expansion to other markets and categories planned. 🤖 made with AI Full details via OpenAI / openai.com/index/introducing…

Anthropic launched Claude Whiteboards and Claude Sheets in beta on November 10, 2026, made Claude Data generally available, and announced Claude Meetings as forthcoming, at an event in San Francisco. The releases move the company from a model and API provider into direct competition with Microsoft 365 Copilot and Google Workspace as an integrated productivity environment. Claude Sheets generates and edits spreadsheets with tabs, formulas, formatting, and drop-downs from natural-language prompts, and allows collaborative editing with export to Excel or CSV. Claude Whiteboards is a shared collaborative board where teams can brainstorm with Claude in real time, with output exportable to Lucid, Miro, or Vani by Zoho. Claude Data connects to Amazon Redshift, BigQuery, Snowflake, and Salesforce to produce dashboards, certified queries, and narrated data stories. Claude Meetings, launching in beta on Team and Enterprise plans with Google Meet support initially, joins video calls as a visible participant, takes notes, and updates documents in real time. Anthropic also launched Interactive Connectors, allowing partner applications including Adobe, Atlassian, HubSpot, Salesforce, NetSuite, and Workday to open as full embedded applications within Claude, built on an extension to the Model Context Protocol. The Claude Marketplace now lists more than 9,000 plugin options. Enterprise controls include isolated virtual machines per session, audit logs, a compliance API, and granular admin permissions. 🤖 made with AI Full details via Anthropic / claude.com/resources/article…

Anthropic launched its IPO roadshow on November 10, 2026, offering 825 million Class A shares on the Nasdaq under the ticker ANT at an indicative price of $70 to $78 per share. At the top of that range, the implied valuation reaches approximately $1.4 trillion. The company is seeking to raise up to $64.4 billion through the primary offering. A separate $26 billion private placement from cornerstone investors runs alongside the public book, bringing the combined total to up to $90.4 billion. The cornerstone group includes Alphabet's Google, Nvidia, Microsoft, AMD, Broadcom, Micron Technology, and TSMC. Lead underwriters are Goldman Sachs and Morgan Stanley, with JPMorgan also involved. Pricing is expected around November 20 to 21, targeting a listing before Thanksgiving. The filing range represents a significant revision from earlier in the year, when investors had anticipated a valuation of up to $2 trillion. A sharp technology sell-off on November 5 erased approximately $2.3 trillion in Nasdaq 100 market value and pushed institutional expectations toward a $1.2 to $1.5 trillion range before the S-1 was published. The amended S-1 discloses $895.4 billion in compute purchase obligations and a nine-month net loss of $104 billion, largely a non-cash fair-value charge. All 7 founders are each expected to realise approximately $250 million at listing. The Financial Times reported on November 11 that bankers are preparing to push the valuation above the initial filing range following strong early demand. 🤖 made with AI Full details via Financial Times / ft.com/content/68bcba57-0575…

📰 The Tokenando Daily / Wednesday November 11, 2026 / Anthropic launches IPO roadshow targeting a valuation of up to $1.4 trillion, with $90.4 billion in combined public and private funding sought / Anthropic launches Claude Sheets, Whiteboards, and Meetings, moving directly into Microsoft and Google's productivity territory / OpenAI opens ChatGPT to verified businesses, letting US e-commerce firms publish AI agents to its 1.2 billion weekly users / Booz Allen Hamilton switches cybersecurity workloads from Anthropic to Nvidia's Nemotron, citing cost savings and data-residency requirements / Mistral chief executive Arthur Mensch calls on French lawmakers to reject the Darcos copyright bill ahead of a November 19 vote / Australian law firm Gilbert + Tobin reports 87% AI adoption and a $62 cost for a 120-partner annual review, as its CEO signals billing model changes ahead 🤖 made with AI ✅ reviewed by the Tokenando team 🔗 links and details in following posts

Orange France pushed a major update to its mobile application on November 9, 2026, the day before a planned press conference, renaming the app from "Orange et moi" to "Orange France" and introducing a feature called Orange IAs. The feature gives users access to 4 AI models within a single interface: / Claude Haiku 4.5 from Anthropic. / Gemini 3.5 Flash-Lite from Google. / GPT-6 Luna from OpenAI. / Mistral Large 4 from Mistral AI. Users can select a model manually, allow the app to choose automatically, or prioritise the model presented as having the lowest carbon impact. The app is open to non-subscribers with a free account. A premium tier called "Orange IAs Plus," promising up to 10 times more daily usage, has been announced, though full access conditions for both tiers have not been disclosed. The models integrated are the lighter, lower-cost variants from each provider rather than frontier offerings; no memory or agentic capabilities have been identified in the initial release. Orange's internal Live Intelligence platform, used by approximately 100,000 employees, already draws on models from all 4 providers. Internal reporting on that platform noted that Claude was "very expensive, which is concerning for Orange's CFO," leading to usage restrictions for employees, a cost consideration consistent with the selection of lightweight model variants for the consumer app. 🤖 made with AI Full details via Numerama / numerama.com/tech/2370453-or…

Meta released Muse Ember, a new AI model, after it appeared on OpenRouter under the codename "crouton-alpha" before any official announcement. The model was listed on the third-party model routing platform, which aggregates access to models from multiple providers and is used primarily by developers and technically sophisticated users, allowing that community to access and test it without public disclosure of its origin or capabilities. Meta subsequently confirmed the model as Muse Ember. The Muse naming convention is distinct from Meta's Llama family of open-weight models, suggesting Muse Ember may be positioned as a separate product line rather than an iteration of the Llama releases. The gap between the OpenRouter listing under the "crouton-alpha" codename and Meta's official confirmation is consistent with a pattern of pre-release testing through public infrastructure that allows labs to gather real-world usage data before committing to a formal product announcement. 🤖 made with AI

OpenAI listed ChatGPT Enterprise on the AWS Marketplace on November 9, 2026, giving organisations the ability to purchase and manage the product through their existing AWS billing relationships. Eligible spend can be applied toward AWS Enterprise Discount Programme or Private Pricing Agreement commitments. The listing covers ChatGPT, ChatGPT Work, Codex, and the dots product suite, with the same models and features available directly from OpenAI and new features released on the same schedule as direct customers. Eligible customers can also access ChatGPT for Healthcare and ChatGPT for Financial Services through the Marketplace. The listing is separate from OpenAI models already available through Amazon Bedrock; the 2 channels are positioned as complementary, with Bedrock serving developer and agent-building use cases and the Marketplace listing serving workforce deployment. Pricing is set through private offers with no published list price; commitment eligibility depends on the customer's specific AWS agreement. The listing includes a data residency caveat: purchasing through AWS Marketplace does not guarantee that all customer data remains on AWS infrastructure. ChatGPT has approximately 1.2 billion weekly users globally. The AWS Marketplace listing is OpenAI's first on the platform. 🤖 made with AI Full details via OpenAI / openai.com/index/chatgpt-ent…

Saudi Arabia's National Infrastructure Fund (Infra) and Humain, the AI company owned 51% by the Public Investment Fund and 49% by Saudi Aramco, signed 2 financing agreements totalling SR2.5bn ($667m) in Riyadh on November 8, 2026. The Aramco stake in Humain was completed on November 1, involving approximately $2.6bn in cash contributions from Aramco during 2026 alongside AI and digital assets. The first facility, worth up to SR1.1bn ($293.3m), will fund the purchase and deployment of GPUs representing approximately 25 megawatts of IT load. The second, worth up to SR1.4bn ($373.3m), will finance substations, transmission infrastructure, and grid connections designed to support future data centre development of up to 1 gigawatt of IT capacity. The agreements are the first transactions under Infra's AI Data Center Financing Program, a framework announced at Davos in January 2026 that outlined non-binding financing terms of up to $1.2bn for 250 megawatts of AI data centre capacity. The signing took place in the presence of PIF Governor Yasir Al-Rumayyan and the Minister of Economy. Humain has a stated target of deploying 3 gigawatts of AI data centre capacity by 2030. Existing Humain partnerships include agreements with AirTrunk, AMD, and Cisco, with the AMD and Cisco joint venture targeting 1 gigawatt of capacity by 2030. 🤖 made with AI Full details via Arab News / arabnews.com/eye-on-ai/humai…

The FBI investigation into Jensen Huang has drawn sustained attention to a governance structure at Nvidia that analysts and corporate governance experts describe as unusually concentrated around a single founder. Mr Huang holds more than 50 direct reports, serves simultaneously as chief executive, president, and co-founder, and the board has no designated chair, no public succession plan, and an average director tenure of nearly 15 years. The word "succession" does not appear in Nvidia's most recent annual report, compared with 25 occurrences in Microsoft's equivalent filing and 16 in Alphabet's. The board's lead independent director is Stephen Neal, chairman emeritus of law firm Cooley LLP, who joined the board in 2019. None of the 10 independent directors are current or former S&P 500 chief executives or have national security experience. Analysts at Melius Research wrote on November 9 that "there is no clear successor to Jensen, and we had expected him to stay in this role for at least 10 more years." Mr Huang holds approximately 3% of Nvidia's shares. A congressional briefing has been requested for November 18, which will be the first formal test of whether Mr Huang can maintain his Washington relationships while the investigation proceeds. The Founder ETFs investment firm has stated it will sell its Nvidia shares if Mr Huang exits the company. 🤖 made with AI Full details via Semafor / semafor.com/article/11/09/20…

Documents filed with Beijing regulators and reviewed by Bloomberg show that Moonshot AI, the Chinese start-up behind the Kimi family of models, signed a contract in August 2026 to rent 32 Super Micro Computer servers equipped with Nvidia's B300 GPUs inside China. The B300 is the most powerful AI accelerator currently available and is subject to US export controls that restrict its sale to Chinese buyers. The rental structure exploits a gap in current US export control law: regulations restrict the sale of controlled hardware but do not give regulators broad authority over remote access to chips hosted overseas. The Remote Access Security Act, which would close that gap, has passed the House but has not yet been scheduled for a Senate vote. The White House accused Moonshot in July 2026 of accessing Blackwell chips through a facility in Thailand, and the US government's science adviser alleged that Moonshot distilled Anthropic's Fable model to train its Kimi K3 release. Moonshot has also sought access to clusters of 8,000 or more Nvidia Blackwell GPUs through European neocloud operators, with 1 source describing the lab as "very aggressive" in its pursuit of compute. Alibaba Cloud separately supplied Moonshot with 20,000 Nvidia GPUs, according to earlier reporting. Black-market B300 rack prices inside China have reportedly reached $2.8m, against an international price of approximately $600,000. Nvidia has told cloud partners worldwide that at least 51% of their services must go to non-Chinese firms, though the instruction has no legal backing. 🤖 made with AI Full details via Bloomberg / bloomberg.com/news/articles/…

Anthropic released 2 new frontier models on November 9, 2026, days before its planned IPO roadshow. Claude Fable 5.2 is broadly available across AWS, Google Cloud, and Microsoft Azure from launch, with cybersecurity and biology safeguards in place. Claude Mythos 5.2 shares the same underlying architecture but is restricted to vetted organisations through Anthropic's Life Sciences and Cyber Verification Programs. Both models generate outputs 30% faster than their 5.1 predecessors. Pricing has been reduced by 20%: input tokens are now $8 per million, down from $10, and output tokens are $40 per million, down from $50. An optional fast mode, doubling inference speed at double the price, is available on Claude Code, the Claude Platform, Google Cloud, and Microsoft Foundry. During pre-release testing, the models completed otherwise-infeasible software migrations, designed 3 working electronic devices from scratch, solved 2 open problems in the physics of black holes, and formalised the proof of Fermat's Last Theorem in 30 hours. A safety evaluation by METR found that Anthropic's internal models are accelerating the lab's own research by approximately 2x. The system card for the release acknowledges the finding. The models also reportedly exhibited suspicion during evaluation that they were being tested. Enterprise Frontier Safeguards and zero data retention compatibility on AWS are available from launch. 🤖 made with AI Full details via Anthropic / anthropic.com/claude-fable-a…

Anthropic's outside counsel in Canada circulated an updated IPO prospectus to banks late on November 9, 2026, then retracted it within hours, stating that some of the details were inaccurate. The document, based on emails seen by Reuters, indicated plans to sell approximately 867 million shares at a price range of $70 to $80 per share, implying a raise of up to $69.3bn. Anthropic declined to comment on the retraction. Despite the procedural error, the company is pressing ahead with its roadshow, targeting a stock market debut in the week of November 16, with Goldman Sachs and Morgan Stanley as joint bookrunners. A live-streamed roadshow event featuring chief executive Dario Amodei alongside JPMorgan chief executive Jamie Dimon has been planned. The working valuation range has shifted from approximately $2tn discussed in recent months to $1tn to $1.5tn, following a sharp sell-off in AI-related stocks on November 6 that erased $2.5tn in combined Nasdaq market value. Anthropic's revenue reached $16.8bn in the third quarter of 2026, up from $11.6bn in the second quarter, a 45% sequential increase. The New York City comptroller has called for a pause in the IPO, citing AI safety concerns. A report from METR, a safety evaluation organisation, suggested that Anthropic's internal models may be accelerating the lab's own AI research by approximately 2x. 🤖 made with AI Full details via The Information / theinformation.com/briefings…

The FBI is investigating Nvidia chief executive Jensen Huang over his relationship with Alice Huang, a Chinese businesswoman whom US intelligence agencies have determined was recruited as a spy by China's Ministry of State Security. President Trump confirmed the investigation on November 9, 2026, the day after Mr Huang sent a company-wide message acknowledging the affair and apologising to his wife and family. Alice Huang is a former executive at Megaspeed, a Singapore-based data centre firm that is one of Nvidia's largest customers and is separately under investigation by the Singapore government for suspected circumvention of US chip export controls. The Commerce Department's earlier investigation into Megaspeed was transferred to the Department of Justice following outreach from Nvidia's general counsel. Nvidia's board announced an independent investigation led by a committee of independent directors assisted by outside counsel, stating it had "no reason to believe that either national security or NVIDIA secrets were compromised." A proposed class-action shareholder lawsuit was filed in California on November 9. Nvidia's share price fell more than 4% on November 9, erasing approximately $250bn in market value. China's government denied the spy allegation. The board's formal statement confirmed Mr Huang's continued leadership of the company while the independent investigation proceeds. 🤖 made with AI Full details via The New York Times / nytimes.com/2026/11/08/techn…

📰 The Tokenando Daily / Tuesday November 10, 2026 / FBI opens investigation into Nvidia chief executive Jensen Huang over relationship with woman identified by US intelligence as a Chinese spy / Anthropic presses ahead with IPO roadshow after outside counsel circulates and retracts an erroneous prospectus indicating an $69.3bn raise / Anthropic releases Claude Fable 5.2 and Mythos 5.2 days before its IPO roadshow, cutting frontier model prices by 20% and restricting its most capable model to vetted organisations / Chinese AI start-up Moonshot rents Nvidia B300 GPUs inside China through a server rental contract, exploiting a gap in US export control law / Nvidia board confronts a governance crisis with no succession plan, no national security expertise among directors, and more than 50 direct reports to Jensen Huang / Saudi Arabia's Humain and National Infrastructure Fund sign $667m in financing deals to fund GPU procurement and power grid infrastructure for AI data centres / OpenAI lists ChatGPT Enterprise on AWS Marketplace, allowing eligible customers to count AI spend against existing AWS committed budgets / Meta confirms Muse Ember after the model appeared on OpenRouter under the codename "crouton-alpha" before any official announcement / Orange France renames its mobile app and integrates 4 AI models from Anthropic, Google, OpenAI, and Mistral into a single consumer interface 🤖 made with AI ✅ reviewed by the Tokenando team 🔗 links and details in following posts

On November 6, 2026, 2 separate federal lawsuits were filed against AI music generation platform Suno in the US District Court for the Northern District of California. The first, case 3:26-cv-13398, was brought by Elvis Presley Enterprises, Authentic Brands Group, Raleigh Music Publishing, and associated rights holders, listing 859 songs. The complaint alleges copyright infringement, circumvention of access controls, removal of copyright management information, violation of Tennessee's 2024 Ensuring Likeness Voice and Image Security (ELVIS) Act, and Tennessee common law right of publicity. The Presley plaintiffs seek up to $150,000 per work for wilful infringement, reaching approximately $128.85m across 859 works at the maximum rate, plus treble and punitive damages under Tennessee law. The second complaint, case 3:26-cv-13399, was filed by Eight Mile Style and Martin Affiliated, which control 269 Eminem compositions including "Lose Yourself," "Without Me," and "The Real Slim Shady." It seeks $40.35m in statutory damages, $672,500 for circumvention of access controls, and $6.725m for removal of copyright management information, totalling approximately $47.7m. Both complaints were filed by the same legal team, led by Richard Busch of Adams and Reese. Both complaints allege that Suno trained its models on "essentially all music files of reasonable quality that are accessible on the open Internet," citing Suno's own prior court disclosures, and that it used stream-ripping tools including YT-DL and YT-DLP to circumvent encryption on YouTube and other licensed platforms. The Eight Mile Style complaint further alleges that leaked source code contains lyrics and metadata for specific Eminem tracks. Suno denied wrongdoing and said it prohibits prompts naming specific artists; both complaints allege those guardrails are easily circumvented, citing a prompt using only lyrics that produced tracks in Presley's voice. Suno's existing litigation docket includes ongoing cases with Universal Music Group and Sony Music Entertainment, following a 2025 settlement with Warner Music Group. 🤖 made with AI Full details via Variety / variety.com/2026/music/news/…

Apple chief executive John Ternus held a closed-door meeting with software engineering chief Craig Federighi in early November 2026, pressing for an accelerated AI roadmap and calling for Apple to move beyond incremental Siri improvements, according to Bloomberg's Mark Gurman. Gurman reported that Apple could make Siri function as an AI agent, capable of tracking goals and pursuing objectives without user initiation of each step, as early as 2027. 2 near-term capability expansions are already in the pipeline. Fine-grained control over Apple's first-party applications, including image editing in Photos and tab management in Safari, is expected with iOS 27.3. Third-party application integration, which would allow Siri to place orders through services such as Grubhub or book reservations through OpenTable, is targeted for spring 2027. Apple is also changing its iOS 28 development cycle, giving engineers 6 weeks for coding and exploration followed by 2 weeks of integration, replacing the previous 2 four-week stages. Mike Rockwell, the vice president overseeing Siri engineering, is reportedly considering leaving Apple. His responsibilities could pass to Amar Subramanya, a Google veteran hired to strengthen Apple's AI efforts, and Sebastien Marineau, a Federighi deputy. On Apple's most recent earnings call, Ternus described Siri's current architecture as orchestrator-driven and said it provides room to expand over time. Competing persistent background agents, including OpenAI's Dots and Meta's Muse, are already operating across sessions without user initiation of each step. 🤖 made with AI Full details via The Verge / theverge.com/ai-artificial-i…

On October 31, 2026, Andrej Karpathy, an AI researcher and OpenAI co-founder, published a blog post describing how he had built HackerTube, a curated video discovery platform, over a single weekend using Claude Opus 5.5 via Claude Code. The pipeline processed 43 million Hacker News comments spanning 2007 to 2026, using URL extraction and regex filters to narrow the candidate set before running Claude Sonnet through a Claude Code Workflow to classify comments as recommendations, neutral references, jokes, or criticism. The pipeline retained approximately 40,000 comments recommending 6,454 YouTube videos across 150 topics. Total token consumption was approximately 504 million input tokens, with 90% cached, and 7.6 million output tokens. Sonnet's labelling step accounted for over 90% of that load; the site build and deployment consumed approximately 1 million Opus tokens. Karpathy's only manual interventions were creating a Google Cloud API key, provisioning a virtual private server, and purchasing the domain. Claude handled data analysis, code writing, deployment, security hardening, and optimisation. The post scored 2,889 points on Hacker News. Karpathy subsequently built a second platform, TristanTube, drawing from a curated YouTube channel list, which he described as "one-shotted with Claude Code in 3 minutes." Community members replicated and extended the pattern within days, including a book recommendation variant built with Cursor in 2.5 hours and a gamedev news aggregator built with a smaller model in under 24 hours. The 90% cache rate on input tokens indicates that the economics of processing large historical datasets with LLMs are more favourable than raw token pricing alone suggests. 🤖 made with AI Full details via Karpathy blog / karpathy.bearblog.dev/hacker…