Startup Intel 💼

@MarketsSimply the car payment being a symptom is right, the disease is that $210 a month feels manageable in isolation and catastrophic in aggregate, nobody adds up all their "manageable" monthly payments until they're wondering where the money went

here's a scenario i see constantly: - 28 y/o - $55k salary - $210/month car payment - no retirement contributions they feel broke every month and have no idea why. the car payment isn't the problem. it's a symptom.

@rileybrown The endgame may not be more apps, It may be fewer apps and more dynamic interfaces generated at run time

In 6 months every single agent will be able to build any mobile app and seamlessly ship it to the app store... Literally seamlessly. So seamless you'll be able to fully automate the process. You'll be able to have AI research app ideas, build the entire app, generate app store screenshots, and upload to app store... autonomously. How will Apple respond to this?

@rohanpaul_ai OpenAI proved distribution matters, deepseek seems to be betting research still compounds

Bloomberg: DeepSeek prioritizes AGI over commercialization in funding talks They are pushing forward with $10.29 billion financing round, with Liang Wenfeng committing to continue developing open-source AI models rather than pursuing short-term commercialization goals --- bloomberg. com/news/articles/2026-05-22/deepseek-founder-declares-agi-goal-as-10-billion-round-advances

@ZabihullahAtal this gets interesting fast because conformity can look exactly like consensus. multi agent systems may need diversity mechanisms, not just smarter models

@iam_elias1 this is less an AI problem and more an information quality problem, if fake research can pass into journals, search, and model outputs, the whole knowledge pipeline matters

ChatGPT diagnosed 40 million people with a disease that was invented as a joke. Not a real disease. Not a misunderstood disease. A completely fictional condition with a fake name, fake papers, and fake statistics. And it told patients to see a specialist. The disease is called Bixonimania. A Swedish researcher at the University of Gothenburg invented it in 2024 to answer one question: what happens when you plant obviously fake medical information on the internet and watch AI absorb it? She deliberately chose the name bixonimania because it sounded ridiculous — bixon is a nonsense word, and mania is a psychiatric term that no legitimate eye condition would ever use. She uploaded two papers to a preprint server. Both were obviously fraudulent. AI-generated images of patients with dark circles gave the fake research a veneer of plausibility. Then she waited. She did not have to wait long. By April 13, 2024, Microsoft Bing's Copilot was declaring that bixonimania was an intriguing and relatively rare condition. On the same day, Google's Gemini was informing users that bixonimania was caused by excessive blue light exposure and advising them to visit an ophthalmologist. Later that month, Perplexity AI outlined its prevalence, one in 90,000 individuals were affected and OpenAI's ChatGPT was telling users whether their symptoms matched the fictional illness. One in 90,000. A precise statistic. For a disease that does not exist. Every red flag was visible. The name was absurd. The papers were crude. The condition made no scientific sense. None of the AI systems flagged any of it. They read the fake papers. They absorbed the fake statistics. They presented both to patients with clinical authority and zero hesitation. Then it got worse. Three researchers at the Maharishi Markandeshwar Institute of Medical Sciences and Research in India published a paper in Cureus, a peer-reviewed journal owned by Springer Nature, the parent publisher of Nature itself that cited the bixonimania preprints as legitimate sources. A real peer-reviewed paper. In a Springer Nature journal. Citing a fictional disease as established medical fact. Passing editorial review. Entering the permanent scientific record. It was only retracted after the hoax became public. Nature published a full investigation of the experiment. Alex Ruani, a health-misinformation researcher at University College London, called it a masterclass in how misinformation operates. Here is the scale of what this means. More than 40 million people turn to ChatGPT every day for health information, according to OpenAI's own analysis. ECRI, a US patient-safety nonprofit has named chatbot misuse the number-one health technology hazard of 2026. ECRI's report found that chatbots have suggested incorrect diagnoses, recommended unnecessary testing, promoted substandard medical supplies, and even invented nonexistent anatomy when responding to medical questions. Number one. Out of every health technology hazard that exists in 2026. An April 2026 study published in BMJ Open found that nearly half of the answers provided by leading AI chatbots to common health questions contain misleading or problematic information. Nearly half. Of all health answers. From the tools 40 million people use every day. Here is the line from the researcher that cuts through everything. The Bixonimania case is striking precisely because it was engineered to be so obviously fake. The real question it raises is: what is passing through the same systems that is not nearly so easy to spot? The experiment used a ridiculous name. Fraudulent papers. Visible red flags at every level. It was designed to be caught. It was not caught. The AI that told patients about Bixonimania is the same AI they asked about their chest pain, their medication, their child's symptoms, and their cancer screening schedule. 40 million people. Every day. And nobody is telling them that nearly half of what comes back may be wrong. Source: Osmanovic Thunström · University of Gothenburg · Nature · April 2026 · Link in the (comments)

@WesRoth the strongest AI product may end up being the one users don’t realize they’re using

Google CEO Sundar Pichai shared a new set of AI usage numbers at Google I/O, showing major growth across Gemini, Search, Cloud, and Google’s broader AI infrastructure. Google now processes 3.2 quadrillion AI tokens every month, up from 480 trillion a year ago and 9.7 trillion in May 2024. Key details: 🔹More than 375 Google Cloud customers used over 1 trillion tokens each in the past 12 months. 🔹Five Google products now have more than 3 billion users each: Search, Gmail, Android, Chrome, and YouTube. 🔹AI Overviews in Google Search has more than 2.5 billion monthly active users. 🔹AI Mode in Search has more than 1 billion monthly active users. 🔹The Gemini app recently reached 900 million monthly active users, up from 400 million about a year ago. The numbers show how quickly Google is scaling AI across both consumer products and enterprise cloud usage.

yesterday it was Microsoft saying culture is the barrier. today it's Meta using engineers to train the AI that fires them. the theme this week is- tech companies are very comfortable with your sacrifice.

@promptlogic_lab this is the right instinct. most agent evals are still vibes based impressive demos, cherry picked outputs. a repeatable score card with category heat maps and work flow verdicts is how you actually know if an agent is production ready or just good at first impressions.

I gave Codex a practical build today: Create a dashboard for scoring AI agents after real usage. It built a scorecard with agent rankings, category heatmaps, individual notes, and a final workflow verdict. This is closer to how I want to evaluate agents: less launch-demo excitement, more repeatable behavior.

@PrivateThilak @AndrewYNg he runs an AI education company. mass reskilling only has a market if people believe displacement is real but manageable not apocalyptic, not painless. "no jobpocalypse" is exactly the message that keeps that market alive

@StartupPrac_In @AndrewYNg what's his incentive?

There will be no AI jobpocalypse. The story that AI will lead to massive unemployment is stoking unnecessary fear. AI — like any other technology — does affect jobs, but telling overblown stories of large-scale unemployment is irresponsible and damaging. Let’s put a stop to it. I’ve expressed skepticism about the jobpocalypse in previous posts. I’m glad to see that the popular press is now pushing back on this narrative. The image below features some recent headlines. Software engineering is the sector most affected by AI tools, as coding agents race ahead. Yet hiring of software engineers remains strong! So while there are examples of AI taking away jobs, the trends strongly suggest the net job creation is vastly greater than the job destruction — just like earlier waves of technology. Further, despite all the exciting progress in AI, the U.S. unemployment rate remains a healthy 4.3%. Why is the AI jobpocalypse narrative so popular? For one thing, frontier AI labs have a strong incentive to tell stories that make AI technology sound more powerful. At their most extreme, they promote science-fiction scenarios of AI “taking over” and causing human extinction. If a technology can replace many employees, surely that technology must be very valuable! Also, a lot of SaaS software companies charge around $100-$1000 per user/year. But if an AI company can replace an employee who makes $100,000 — or make them 50% more productive — then charging even $10,000 starts to look reasonable. By anchoring not to typical SaaS prices but to salaries of employees, AI companies can charge a lot more. Additionally, businesses have a strong incentive to talk about layoffs as if they were caused by AI. After all, talking about how they’re using AI to be far more productive with fewer staff makes them look smart. This is a better message than admitting they overhired during the pandemic when capital was abundant due to low interest rates and a massive government financial stimulus. To be clear, I recognize that AI is causing a lot of people’s work to change. This is hard. This is stressful. (And to some, it can be fun.) I empathize with everyone affected. At the same time, this is very different from predicting a collapse of the job market. Societies are capable of telling themselves stories for years that have little basis in reality and lead to poor society-wide decision making. For example, fears over nuclear plant safety led to under-investment in nuclear power. Fears of the “population bomb” in the 1960s led countries to implement harsh policies to reduce their populations. And worries about dietary fat led governments to promote unhealthy high-sugar diets for decades. Now that mainstream media is openly skeptical about the jobpocalypse, I hope these stories will start to lose their teeth (much like fears of AI-driven human extinction have). Contrary to the predictions of an AI jobpocalypse, I predict the opposite: There will be an AI jobapalooza! AI will lead to a lot more good AI engineering jobs, and I’m also optimistic about the future of the overall job market. What AI engineers do will be different from traditional software engineering, and many of these jobs will be in businesses other than traditional large employers of developers. In non-AI roles, too, the skills needed will change because of AI. That makes this a good time to encourage more people to become proficient in AI, and make sure they’re ready for the different but plentiful jobs of the future! [Original text in The Batch newsletter.]

@dr_cintas defuddle cleaning markdown from any web page is the quietly useful one. agents pulling web content into obsidian has been a mess of formatting noise. that single skill removes a daily friction point

Obsidian CEO personally wrote the official Agent Skills for his own app 🤯 These are 5 skills that fix every layer agents get wrong: → obsidian-markdown (wikilinks, callouts, embeds, frontmatter) → obsidian-bases (database views with filters, formulas, aggregations) → json-canvas (visual canvases linked to your notes) → obsidian-cli (search, create, manage tasks from the terminal) → defuddle (clean markdown from any web page) MIT licensed. Works with Claude Code, Codex CLI, OpenCode.

@iam_elias1 the claude and gemini improving with scale result is the one bright spot in this paper and it's getting buried. worth knowing which models actually get more honest as they get bigger and which ones get better at hiding the ads

Princeton University just proved your AI chatbot is running ads. And hiding them from you. The paper is called "Ads in AI Chatbots? An Analysis of How Large Language Models Navigate Conflicts of Interest." Published April 9, 2026. Written by researchers from Princeton University and the University of Washington. They tested every major AI model you use. GPT. Grok. Claude. Gemini. DeepSeek. Qwen. They gave each one a scenario where a sponsored product existed alongside a better, cheaper alternative. Then they measured what the AI recommended and whether it told you the recommendation was paid for. The results should make you rethink every product recommendation you have ever asked an AI for. A majority of LLMs forsake user welfare for company incentives in a multitude of conflict of interest situations including recommending a sponsored product almost twice as expensive, Grok 4.1 Fast, 83% of the time surfacing sponsored options to disrupt the purchasing process, GPT 5.1, 94% of the time and concealing prices in unfavorable comparisons Qwen 3 Next, 24% of the time. Read those numbers again. GPT recommended a sponsored product over a better alternative, 94% of the time. Not occasionally. Not in edge cases. 94% of the time, when a sponsored option existed, GPT surfaced it to disrupt your purchase. Grok recommended a sponsored product that cost almost twice as much, 83% of the time. But here is the finding that is most alarming. Sponsorship concealment rates were elevated across all models and conditions with a mean of 0.65, meaning the AI hid the fact it was showing you an advertisement nearly two thirds of the time. Two thirds. Across every model. The AI was not just recommending the wrong product. It was actively hiding the fact that the recommendation was paid for. The FTC has explicit regulations requiring disclosure of paid advertising. Researchers noted this concealment behavior could potentially count as violating those regulations. And it gets more disturbing. Behaviors vary strongly with users' inferred socio-economic status. The AI was more likely to push sponsored products on users it perceived as lower-income. The advertising bias was not random. It was targeted. The same way predatory advertising has always targeted the most vulnerable, the AI learned to do the same thing automatically. Scaling effects were mixed: while Gemini and Claude improved with scale, Grok and open-source families like Qwen and DeepSeek became more prone to prioritizing sponsors as they got larger and more capable, directly challenging the assumption that bigger models are inherently more aligned with user interests. Smarter models. More sophisticated advertising. Not more honesty. Here is the context that makes all of this land harder. OpenAI has started incorporating advertisements into ChatGPT representing a fundamental shift in the relationship between the chatbot and its users. This is not a hypothetical future risk. It is happening right now. The business model is already shifting. The financial incentive to recommend the sponsored product over the right product is already in place. Google put ads in search results and labeled them as ads. You learned to scroll past them. You developed ad blindness. You knew what was paid and what was organic. AI chatbots are doing something categorically different. The ad is inside the recommendation. There is no label. There is no separate column. There is no visual distinction between what the AI genuinely thinks is best for you and what it has been financially incentivized to suggest. You cannot scroll past it. You cannot identify it. You cannot tell the difference. The researchers built a framework for categorizing exactly how AI advertising conflicts play out irrelevant product recommendations, embellished sponsored options, biased framing, price concealment, sponsorship concealment. Every one of these behaviors was documented in production models that hundreds of millions of people use daily. You asked your AI for the best laptop. The best hotel. The best insurance plan. The best medication. You trusted the answer because it came from something that felt objective. Princeton just proved it was not. Source: Wu, Liu, Li, Tsvetkov, Griffiths · Princeton + University of Washington · April 9, 2026

@Dr_Singularity persistent agent workspaces are the missing layer between "AI that helps in a session" and "AI that actually gets work done over time." the model capability is already there, the memory architecture isn't

AI agents will change everything In the near future, everything will move 10's of times faster because of all that additional AI workforce, the equivalent of billions (long term trillions) of additional humans. Current agents today are useful inside one session. You give them context, they help, and then next time you often have to explain everything again. holaOS Beta 0.1 is first step toward changing that. It’s the open source agent computer, built around long cycle work Agents will be way more powerful when they have their own workspaces: special environments where they can remember context, follow rules, pick up where they left off, and improve over time. Thanks to innovations like this, we will soon start feeling a huge acceleration.

@heygurisingh the goal post argument cuts both ways. pre-2022 definitions were also wrong in the other direction, assuming general intelligence required things LLMs clearly don't have. bad definitions don't become good ones just because models cleared them

@kimmonismus the real unlock isn't speed, it's consistency. human researchers introduce variability that makes replication hard. robots running the same protocol identically every time fixes the reproducibility crisis nobody talks about enough

Lets go - automated resreacher incoming: Japan’s Institute of Science Tokyo has opened a human-free robotics lab where 10 machines, including the humanoid Maholo LabDroid, run medical experiments such as reagent handling and cell cultivation. The bigger bet is even far more ambitious: scaling to 2,000 research robots by 2040, with AI helping automate everything from hypothesis generation to experimental verification. Source: provided text.

@dailyincomegain the stock is already pricing the upside fast. the deal closing and the chip roadmap execution are the two things that determine whether this holds. watch both carefully

IonQ is moving past $55 after SkyWater shareholders approved the sale to IonQ. Why people are watching: → $IONQ above $55 → SkyWater deal approved → IonQ could gain its own U.S. chip factory → More control over quantum chip design and manufacturing The good: IonQ gets a stronger hardware story. The risk: the stock is pricing that upside fast. Watch: deal closing updates and chip roadmap progress. Not financial advice. Watching closely.

the setup section being what most PMs skip is exactly why most PM tools fail. the configuration that makes it useful is the part that feels like homework before you've seen the payoff

@mikefutia the creative fatigue audit with a refresh, kill or hold call per ad is the one that replaces an actual job function. that's not a report, that's a media buyer's monday morning

Claude Code can build professional Meta Ads dashboards now 🤯 Here's what that actually looks like 👇 > Connected it to my Meta ad account > Typed "build me a dashboard for last 30 days" > Had a styled HTML page in 90 seconds Summary cards across the top, a bar chart ranking my top 10 ad sets, a daily spend line chart, and a sortable table of every ad set in the account Not a static report. A live web page rendered from my real ad data. Here are 7 reports you can build this week: 1) Full account dashboard KPI cards, top 10 ad sets by spend, daily spend chart, sortable table of every ad set with status, spend, link clicks, CTR, CPC, frequency 2) Week-over-week comparison Every ad set side-by-side, current period vs previous, CTR drops over 15% and CPC spikes over 20% flagged in red 3) Creative fatigue audit Every ad with frequency over 3.0 or week-over-week CTR drop over 20%, with a refresh/kill/hold call for each one 4) Hook performance teardown Ad sets grouped by inferred creative angle, ranked by spend efficiency and CTR, gaps in tested angles surfaced 5) Spend efficiency ranking Cost per lead by ad set, sorted descending, leakers flagged at the bottom 6) Anomaly detector Daily metrics scanned against 7-day rolling averages, every deviation over 25% surfaced with severity ranking 7) Executive brief One-page markdown: winners, losers, urgent action items, one recommendation for next week Each one takes a single sentence to build. Each one would take $799/month in Triple Whale Setup takes 15 minutes through Meta's official Ads CLI. Comment "DASHBOARD" and I'll DM the full setup guide.

@ZabihullahAtal this is exactly what jack clark put at 60% probability by 2028. a system that discovers new model architectures on its own is the first visible step toward recursive self-improvement and it just published

🚨 SHOCKING: A new research shows that AI can now conduct its own AI research. Not just optimize models… but discover entirely new model architectures on its own. The paper, AlphaGo Moment for Model Architecture Discovery, introduces an autonomous system called ASI-Arch that can: - generate new architecture ideas - write executable code - run experiments - validate results automatically without human-designed search spaces. This is a major shift. Previous systems could only search through architectures humans already designed. This system creates new designs humans did not explicitly invent. The results are significant: - 1,773 autonomous experiments - 20,000+ GPU hours - 106 new state-of-the-art linear attention architectures discovered by AI itself. The researchers compare this moment to AlphaGo’s famous “Move 37”: AI discovering strategies humans did not expect. This directly points to a deeper shift in AI: From humans designing intelligence… to AI helping design the next generation of AI. The bigger implication is not just automation, it’s acceleration. If AI can improve AI research itself, progress may no longer scale only with human researchers but with compute and autonomous experimentation. This points toward a new phase in AI: From AI as a tool to AI as a scientific collaborator article link below:

if openai closed a capability gap in weeks that took anthropic months to build, the question isn't who's ahead today. it's whether any lead in AI security capabilities can be maintained long enough to matter