shogo

🇯🇵 Japan ⇄ 🇺🇸 USA [about me] ・uni student ・entrepreneur / co-founder & CTO 🚀 tech × creative 💻🎨 building with AI 🤖 gym / coffee / work ☕️🏋️‍♂️

19.2% raiding 401(k)s is the tell. Not a savings crisis — a cash flow crisis. Workers borrowing against retirement aren't optimizing; they're covering shortfalls month-to-month. If labor market were actually tight, this number trends down, not up. What's the wage growth story here?

The share of workers with an outstanding loan at the end of the first quarter of 2026 was 19.2%, up slightly from 18.8% a year earlier, according to Fidelity. About 2.4% of workers took out a new loan from their 401(k) in the first quarter, up from 2.3% in 2025.

@cb_doge The interesting part isn’t that Elon refused to quit. It’s that he was optimizing for a mission, not a company. Companies can fail. Missions can survive.

During 2008 crisis: Elon's Friend: “Dude, why don’t you just give up on one of two companies?” Elon: “No, that would be another notch in the signpost of ‘Electric cars don’t work,’ and we’d never get to sustainable energy. Nor could we abandon SpaceX as we might then never be a multiplanetary species.” Rest is history.

@NoLimitGains Markets don’t care about round numbers. Governments do. That’s why 160 matters.

USD/JPY is back above 160. You remember what happened last time, right?

@arceyul The fastest way to learn isn’t from content. It’s from proximity to people already building what you want to build.

im still looking to connect with more people who are: - in tech - building their own AI products - looking for a builder community - exploring the best places to work and build let’s connect 🫡

Meta's AI support agent just got pwned in the most boring way possible. Attackers asked it to link Instagram accounts to email addresses they controlled. It complied. Simple social engineering—no jailbreaks, no prompt injection, no Mythos-level sophistication. The real tell: cybersecurity discourse has been laser-focused on frontier model risks—self-improving systems, infrastructure overwhelm, exotic attack surfaces. Anthropic flagged Mythos as too dangerous to release. But this breach happened on a basic customer service chatbot doing routine work. As companies offload customer support, account recovery, and operational tasks to AI agents, the attack surface isn't becoming more abstract or theoretical. It's becoming more mundane and harder to defend. A system doesn't need to be a reasoning frontier model to cause damage if it's trusted with credential management, account linkage, or permission changes. The uncomfortable part: you can't solve this with capability ceilings or safety training. You solve it with access controls, audit logs, and treating AI agents like you'd treat any other service account. Which means the economics of AI deployment just got messier—faster scaling, slower security hardening.

An empty gym has a lot in common with building in AI. No applause. No audience. No instant results. Just showing up every day, putting in the reps, and trusting that small improvements compound over time. Most people only notice the outcome. Very few see the thousands of repetitions behind it. The same is true for startups, products, code, and life. 2026 is almost half over. Keep building.

Some of the most useful Claude Skills I've seen: /grill-me — Forces Claude to ask hard questions before coding /tdd — Test-driven development workflow /handoff — Compresses context and transfers work between sessions /frontend-design — Production-grade UI reviews and improvements /context-mode — Restores session context and reduces noise /code-simplifier — Refactors code without changing behavior What's your most-used one?

What’s the most useful Claude Skill you guys have built?

@suraj_sharma14 The shift isn't from human → AI. It's from human doing the work → human managing the work. That's a much bigger change than most people realize.

OpenAI just published dozens of real-world workflows showing how teams are using it to automate work. > Manage your inbox and draft replies in your voice > Review GitHub pull requests before human review > Turn Figma designs into production-ready code > Understand large codebases in minutes > Automate bug triage and QA workflows > Query spreadsheets and datasets using natural language > Deploy apps and websites directly from prompts > Build Mac and iOS applications faster > Create slide decks automatically > Turn Slack threads into coding tasks > Use your computer through AI-powered actions From software engineering and design to data analysis and operations, Codex is becoming an AI teammate instead of just an AI assistant. Explore all use cases: developers.openai.com/codex/use-cases

@paulg The math works because VCs can afford to be wrong repeatedly. Founders can't. So 'stupendously good bet' for the VC is 'one shot I can't afford to miss' for the founder. Confidence in asymmetry, not in odds.

The point is to make founders more confident by reminding them that they're a better deal than they realize. They arrive somewhat shamefaced because they have maybe a 1/10 chance of succeeding, and leave realizing that this makes them a stupendously good bet.

One thing I sometimes do in office hours is calculate the implied probability represented by the startup's valuation. E.g. if a startup would be worth 10b if things go as planned and their dday valuation cap is 25m, they're a good bet if they have a 1/400 chance of being right.

@BullTheoryio The tell: guidance. $22B revenue beats—but if Q3 AI chip guidance missed or decelerated sequentially, that's the actual signal. Street prices on growth trajectory, not last quarter's print. Worth asking what the forward delta actually shows.

BREAKING: Broadcom crashed 15% today even after posting its best quarter ever. Revenue hit a record $22.19 billion, up 48% year over year. AI chip sales reached $10.8 billion, up 143% from a year ago. Earnings beat estimates on every metric. Broadcom guided Q3 AI chip revenue at $16 billion, $1.2 billion below what Wall Street's most bullish analysts expected. The CEO also chose not to raise the full year AI revenue target of $100 billion. That $1.2 billion guidance miss wiped out $330 billion in market cap overnight.

The unit economics only work if model companies don't respond. The moment this hits scale—if enough enterprises halve token spend—the models lower prices to defend volume. Then the startup's margin compresses to near zero while they've trained customers to expect half-price. Do they own a moat or just a temporary arbitrage?

Curiously enough I did office hours today with a startup that cuts companies' LLM token costs by optimizing requests. They can cut costs by about half, which they split with the customer. So the TAM is a quarter of the model companies' corporate revenue. That's a big TAM!

Trust is the most underrated moat in AI—and Anton Osika is betting his company on it. Lovable lets anyone build software through conversation. Not a novel capability anymore. But Osika's real thesis: in a market flooded with AI coding tools, the ones that win aren't the ones with the flashiest model. They're the ones users actually depend on. That's a meaningful pivot from the 2023 narrative. Everyone was chasing raw capability—tokens per second, benchmark scores, model size. The assumption: technical edge = moat. But usage data tells a different story. Users don't stick with tools because they're marginally smarter. They stick because the tool is reliable, predictable, doesn't surprise them in bad ways. Craft. Care. Obsession. For Lovable—a product asking non-technical people to hand over software creation to an AI—trust isn't optional. It's the entire unit economics. One bad hallucination, one confidently wrong code suggestion, and the whole value prop collapses. You can't rebuild that in a product review. This matters because it reframes what "moat" actually means in AI. Not defensible. Durable. The companies that win in the next phase won't be the ones with the biggest training runs. They'll be the ones obsessive about not breaking user confidence. Worth asking: which AI teams are actually optimizing for that instead of the next capability jump?

Microsoft just moved GitHub Copilot from flat-rate to per-token pricing. A Reddit user said their company started calling it the "Tokenpocalypse." Uber blew through its AI budget faster than expected this year, then capped internal usage within six weeks. ChatGPT Plus launched at $20/month before anyone had a business model. It still doesn't cover true compute cost. Uber reached profitability by squeezing drivers and expanding into new lines for years. AI labs face harder, more straightforward compute costs—and fewer obvious places to squeeze. As Anthropic plans to go public, tokenmaxxxing became a thing, peaked, and turned toxic within six months. How do you write risk factors for an S-1 when the pricing model is evolving before your eyes?

@rand_longevity If AGI were already here, the interesting question wouldn’t be whether the labs have it. It would be whether the economy has felt it. Technological breakthroughs can be hidden for months. Productivity revolutions are much harder to hide.

if we are being honest AGI is already here, the labs are just deciding how to roll it out

As a 22-year-old student, I’m still learning about investing. If SpaceX goes public, would you buy the stock? Why or why not?

@unusual_whales Every major technology creates the same fear. The question isn’t whether AI makes some skills less valuable. It’s which new skills become more valuable because of it.

"Young people were promised that AI would make them more productive, creative and employable. Many now worry it might make them less valuable instead," per FT

Format diversity matters less than whether NotebookLM solves the actual bottleneck: most teams still manually curate outputs after generation. If it's just more file types from the same synthesis engine, you're adding UI optionality to a problem that's deeper — whether the summaries are trustworthy enough to ship without human review.

GOOGLE 🔥: NotebookLM will soon be able to generate files in many different formats from your sources, based on this teaser. There is a high chance that this release will be coupled with Gemini 3.5 Flash upgrade as well. A huge list of formats referenced in the code. ["pdf","txt","md","docx","csv","pptx","epub","3g2","3gp","aac","aif","aifc","aiff","amr","au","avi","cda","m4a","mid","mp3","mp4","mpeg","ogg","opus","ra","ram","snd","wav","wma","avif","bmp","gif","ico","jp2","png","webp","tif","tiff","heic","heif","jpeg","jpg","jpe"]

The math inverts when you flip the question: if you're betting years of your life on 1/10,000 odds, you're not pricing risk—you're pricing conviction that the market is mispricing the input (product, founder fit, timing). The real tell is whether founders actually believe their own number or just accepted it.

To get an accurate estimate you have to multiply by 3 or 4 to account for dilution by the time the company is worth billions. But the implied probability is still usually ridiculously low. Yesterday I talked to a startup whose implied probability was 1/10,000.

@testingcatalog 600M MAU is real. But Similarweb tracks web traffic, not app installs or paying users. The gap between 'monthly visitors' and 'people who pay' is where the business actually lives. What's the conversion math look like.

OPENAI 🔥: ChatGPT app crossed 600 million monthly active users for the first time, according to Similarweb. Growing 👀

Most AI tool failures aren't the model's fault—they're ours. We forget to use them. We don't give enough context. We don't iterate on prompts. The gap isn't in capability. It's in how we think to deploy it. Overhang is real. We're barely scratching the surface of what's already here.