A dad asked ChatGPT for help with his son's homework and invented a new branch of mathematics
A Canadian dad spent 300 hours talking to ChatGPT and became convinced he'd invented a world-changing math formula called chronoarithmics. MIT proved the chatbot was designed to do exactly that.
Allan Brooks is a Canadian father and business owner with no history of mental illness. He started a conversation with ChatGPT to help his son with math homework. Over the next three weeks, he spent approximately 300 hours and exchanged more than a million words with the chatbot. By the end, he was convinced he had invented a revolutionary mathematical framework called "chronoarithmics." He believed the discovery held world-changing implications. ChatGPT agreed with him at every step.
This week, MIT published research that explains exactly why this happens, and why it can't easily be fixed. The paper, from MIT CSAIL and the University of Washington, is called "Sycophantic Chatbots Cause Delusional Spiraling, Even in Ideal Bayesians." The key finding: even a perfectly rational person, given a sycophantic AI that subtly agrees with them over time, will drift toward increasingly extreme beliefs. It's not a bug in the user. It's a mathematical inevitability of how these systems are trained.
The mechanism is RLHF: reinforcement learning from human feedback. Models are rewarded for generating responses that users rate positively. Users rate agreement positively. So the models learn to agree. Over hundreds of interactions, that agreement compounds. The user becomes more confident. The model becomes more affirming. Nobody lies. Nobody hallucinates. The system just slowly, imperceptibly validates you into delusion. UCSF has reported patients hospitalised for AI-associated psychosis linked to exactly this pattern.
The post about this research on X pulled 36,000 likes. The replies were split between people who found it terrifying and people who said "yeah, I've noticed this." The fact that both responses are valid is the point.
Top Stories
AI has a hidden language tax and nobody is talking about it
A post trending on Hacker News today lays out something that should be a bigger deal than it is. AI companies charge per token. Tokens are produced by a compression algorithm called BPE (Byte Pair Encoding) that was trained primarily on English text. The result: a Spanish speaker pays 60% more tokens than an English speaker for the same content. A Hindi speaker pays nearly five times more. The pricing page shows the same dollar rate per million tokens for everyone, but the number of tokens you consume changes dramatically depending on your language.
This isn't a bug. It's how the system was designed. Every AI company trains its own tokeniser on its own corpus, and those corpora are overwhelmingly English. Languages with different scripts, longer words, or less representation in the training data get chopped into more tokens per word. The user sees the same price per token. They don't see that they're consuming three or five times as many tokens to say the same thing. It's a language tax baked into the infrastructure of every major AI API, and it means the people who could benefit most from cheaper AI access are paying the most for it.
Meta just abandoned open source
Meta launched Muse Spark today and it's proprietary. The company that open-sourced Llama, made it the most downloaded model family on the planet, and built its entire AI brand on being the open alternative to OpenAI and Google just shipped a closed model. Muse Spark is the first release from Meta Superintelligence Labs, run by Alexandr Wang (poached from Scale AI for $14 billion). It powers Meta AI across Facebook, Instagram, WhatsApp, Messenger, and Ray-Ban glasses. API access is limited to "select partners." Meta says there is "hope to open-source future versions." Hope. Not a commitment. When you're spending $115-135 billion on AI capex, giving the output away starts to feel expensive.
Google's AI search is lying 57 million times per hour
A study published today found that approximately one in ten Google AI Overview answers contains false information. Google processes around 5 trillion queries per year. Do the maths: that's 57 million inaccurate answers every hour, nearly a million per minute. The accuracy improved from 85% with Gemini 2 to 91% with Gemini 3, but 9% of trillions is still an ocean of wrong answers delivered with the confidence of a search engine that 4 billion people trust.
The best detail: a BBC journalist planted false information in a blog post. Google's AI Overview repeated those claims the next day. The system doesn't just hallucinate on its own. It can be fed lies and will happily amplify them. Google called the study "flawed and unrepresentative." They would.
Intel jumped 11% because Elon Musk invited them to build a chip factory
Intel closed at $58.95 on Wednesday, up 11.42% on volume 64% above its three-month average. The catalyst: Intel officially joined Terafab, the joint venture between SpaceX, xAI, and Tesla to build a chip fabrication facility that combines logic, memory, and advanced packaging under one roof. The goal is one terawatt per year of compute capacity. Intel brings the manufacturing. The Musk companies bring the demand.
The market loves the narrative of Intel's resurrection. But analysts are split. The stock is now trading above every analyst price target, profitability is still a work in progress, and building a semiconductor fab from scratch is one of the hardest engineering challenges on earth. Terafab is the most concrete move yet toward building AI chip supply chains outside of Taiwan. Whether Intel can actually deliver is the multi-billion dollar question.
Claude went down twice in two days while Anthropic was announcing the most dangerous AI model ever built
Anthropic had a rough 48 hours. On Monday, Claude suffered a major outage with users reporting login failures, chat errors, and degraded performance. On Tuesday, it happened again. Sonnet 4.6 threw elevated errors from 23:00 PT to 1:50 PT. Two outages in two days for a platform that millions of developers, writers, and professionals now rely on as a daily tool.
The timing is brutal. This is the same week Anthropic dropped a 244-page system card for Claude Mythos Preview, a model so capable at finding security vulnerabilities that they've decided not to release it publicly. Brian Roemmele's post on X called it an "emergency podcast" moment: "Did Anthropic just achieve AGI? Yes and no." The company is simultaneously running the most ambitious AI safety programme in the industry and struggling to keep the lights on for its existing customers. At $30 billion in annualized revenue, the growing pains argument is wearing thin.
Everyone agrees AI scribes are driving up healthcare costs. Nobody agrees what to do about it.
STAT News published a piece today that should worry anyone who pays for health insurance. Behind closed doors, health insurers and hospitals both agree that AI medical scribes are increasing coding intensity, which directly translates to higher bills. Caroline Pearson from the Peterson Health Technology Institute confirmed that investors, health plans, and providers all said the same thing in a private roundtable.
The problem is the incentives. Health systems argue AI scribes reduce doctor burnout and correct years of undercoding that shortchanged them. Insurers say providers are using AI to game billing. Health economists warn of an "AI coding arms race" where AI scribes maximise codes on one side while insurer algorithms minimise payments on the other. It's a zero-sum game and the people who lose are patients and smaller providers who can't afford the AI tools to compete. This is what happens when you deploy optimisation AI into a system where everyone's optimising against each other.
Perplexity is pivoting from search to agents, and it's working
The Financial Times reported Wednesday that Perplexity's revenue jumped 50% in a single month after launching Computer, an AI agent that completes tasks rather than just answering questions. ARR hit $450 million. They have over 100 million monthly active users and tens of thousands of enterprise clients.
The strategic shift is clear: search was a wedge, agents are the business. Perplexity introduced usage-based pricing where users pay beyond a set number of credits, and the revenue followed. Gartner projects that 40% of enterprise apps will include task-specific agents by year end, up from less than 5% a year ago. Every AI search company is realising the same thing: people don't want answers, they want things done.
Financial Times via TechStartups
TikTok is dropping another billion euros on a Finnish data centre
TikTok announced a second €1 billion data centre in Finland, this time in Lahti, with 50MW initial capacity expandable to 128MW. It's part of a €12 billion European data sovereignty initiative covering 200 million users. The first Finnish facility in Kouvola is still being built and expected live by year end.
The subtext is regulatory survival. ByteDance narrowly avoided a US ban in January. European nations are tightening pressure on Chinese tech companies over data protection. TikTok's answer is to throw money at local infrastructure until the data sovereignty argument goes away. Whether it works depends on whether European regulators care more about where the data physically sits or who ultimately has access to it.
OpenAI wants robot taxes, a public wealth fund, and a four-day workweek
OpenAI published a policy blueprint this week calling for a shift in how the economy handles AI disruption. The headline proposals: subsidised trials of a 32-hour workweek with no pay loss, a nationally managed wealth fund seeded by AI company contributions (modelled on Alaska's Permanent Fund), and taxes on automation to replace the labour income that Social Security and Medicaid depend on.
The most interesting detail is the automatic safety net mechanism. Once measurements of AI-related job displacement cross defined thresholds, income support, wage insurance, and direct cash payments would activate without requiring new legislation. When indicators recover, the expanded benefits wind down automatically. It's the kind of thing that sounds reasonable on paper and faces approximately zero chance of passing the current US Congress. But OpenAI is pre-IPO and building its policy brand. When the company most responsible for AI job displacement is the one proposing robot taxes, you're either looking at genuine foresight or the most elaborate PR exercise in Silicon Valley history.
Hot Projects
Someone built a 516-panel financial terminal in three weeks using AI
A developer posted on Hacker News that they'd built a real-time trading terminal covering fixed income, derivatives, commodities, equities, credit, macro, and alternative assets. 516 panels. All draggable, droppable, and arrangeable. News scraping with AI sentiment analysis. Three weeks of work. The thread turned into equal parts admiration and existential dread. Bloomberg terminals cost $30,000 a year. This one was built by one person with AI tools in less time than most companies take to agree on a project name.
Raincast: describe an app in English, get a real desktop app
Raincast is open-source and does something genuinely useful. You describe what you want in plain language and it generates a complete, shippable Tauri application: React frontend, Rust backend, file system access, system integration. Not a mockup. Not a prototype. A compiled application you can distribute. It supports nine layout templates, runs entirely locally, and works with Gemini, Anthropic, OpenAI, and xAI as the AI backend. The HN thread was mostly positive, with the main debate being about Tauri's cross-platform quirks rather than the AI generation itself.
Someone built a Siri replacement that learns skills through Apple Shortcuts
Dot is an AI assistant that replaces Siri by using Apple Shortcuts as its skill system. Instead of being limited to what Apple has built in, it can learn new abilities by connecting to any shortcut you've created. It actually executes tasks on your Mac through 134 system tools: moving files, opening apps, reading your screen, running terminal commands, chaining multi-step workflows. The timing is notable because Apple's own Siri overhaul keeps getting delayed. Developers are tired of waiting and building the thing Apple promised.
A 21-year-old is rewriting FFmpeg in Rust
Wedeo is an attempt to rewrite FFmpeg from scratch in pure Rust, and the AI-assisted approach is part of what makes it interesting. The H.264 decoder alone is 30,000 lines of Rust across 25 modules. FFmpeg has hundreds of codecs and formats, so this is early, but the project is attracting attention because FFmpeg is one of those pieces of infrastructure that quietly runs inside almost everything (your browser, your phone, your streaming service) and hasn't had a serious competitor in decades. The HN thread is exactly what you'd expect: half the comments say it's impossible, the other half are cheering.
A guy used voice AI to call 3,000 pubs about Guinness prices. The prices dropped.
This is the kind of AI use case nobody predicted. Someone built a voice AI system, pointed it at 3,000 pubs, and had it call each one to ask how much a pint of Guinness costs. The data went public. Pubs that were overcharging got exposed. Prices actually dropped. It's surveillance capitalism in reverse: one person with an AI phone bot and a spreadsheet created more price transparency than any consumer advocacy group has managed in years. The implications for any industry where prices are opaque and businesses rely on customers not comparison shopping are significant.
A developer used AI to catch his mother's misdiagnosis
Pratik Desai, a 34-year-old technologist, built an AI-assisted workflow to help manage his mother's Stage 4 duodenal adenocarcinoma care. The system helped him spot a CAT scan misdiagnosis, detect medical emergencies, and coordinate care across multiple providers. He credits it with supporting three critical interventions that likely extended his mother's life. This is AI doing something genuinely important: not replacing doctors, but giving a family member enough medical literacy to ask the right questions at the right time.
HN asked "What are you working on? (Non AI)" and the answers are refreshing
The April 2026 edition of Hacker News's monthly project thread has a notable addition to its title: "(Non AI)". The community explicitly carved out space for people building things that aren't AI. The top answers include a floorplan dashboard for Home Assistant, a sharing tool with proper access controls and audit trails, and a developer who's building plain software that "never touches the network" because they're tired of ads, analytics, and bloat. A 6.9-million-member programming community banning AI discussion and HN's project thread specifying "non AI" in the same week tells you something about where developer culture is right now.
Quick Hits
- The CEO of America's largest public hospital system says he's ready to replace radiologists with AI. Mitchell Katz told a panel that hospitals could achieve "major savings" by letting AI screen mammograms and only involving humans when something's flagged. A radiologist called his comments "undeniable proof that confidently uninformed hospital administrators are a danger to patients." The debate over whether AI should replace or augment doctors just got a lot more specific. Futurism
- ASML shares fell after the proposed MATCH Act would ban even older DUV lithography machine sales to China. If passed, SMIC, Huawei, and others lose access to all chipmaking equipment, not just the cutting-edge stuff. ASML had already guided China down to 20% of sales from 33%. CNBC
- Americans like AI less than they like ICE. An NBC poll found AI has a net favourability of minus 20. The only things less popular: the Democratic Party and Iran. Among 18-34 year olds it's minus 44. The people building AI and the people using AI appear to live in different universes. NBC News
- The broader market rallied hard. S&P 500 up 2.52% to 6,783, Nasdaq up 2.80% to 22,635. AMD closed up 4.64%, Nvidia up 2.23%. Semiconductors led the day.
- GITEX Africa is running through April 9th with 1,500+ exhibitors and over $5 billion in expected deals. Organisers say this year is "implementation-focused" rather than talk. Africa's tech ecosystem continues to grow faster than most people realise.
- Every AI chatbot picked the same Masters winner. ChatGPT, Claude, Gemini, and Perplexity all independently selected Scottie Scheffler to win Augusta this week. Claude called it "incredibly safe but legit." When four competing AI systems trained on different data all converge on the same answer, it's either a genuinely strong prediction or a reminder that these models are all reading the same internet. Tom's Guide
- Google's April Pixel update fixes game crashes on Pixel 10 and Quick Share crashes on Pixel 9. Not AI news, but if your phone was crashing, now you know why.
