Every week, get the latest curated high signals on news, insights, and ideas around technology, science, and business to help you become a better builder and thinker. Please fill out the survey so we can deliver the best content to your inbox. Thanks for reading!
Weekly Find: Reddit is quickly becoming the source of truth for AI models.
This is quite a telling graph because it’s becoming much clearer now what data these LLMs are being trained on. You’re also starting to see these LLMs like Grok (by xAI) that have proprietary access to live feeds from X where real, authentic, and genuine user-generated conversations are becoming ground truth.
Following behind are Wikipedia, YouTube, Google, and even Yelp. This makes sense, since these sources cover long-tail queries that traditional news or academic sources don’t. They also contain opinion, personal experiences, step-by-step instructions, and up-to-date local details, which is exactly what users ask LLMs! We all know there’s too much crap on the web, so having user- generated content matters.
But one thing we need to be aware of is that heavy reliance on user-generated content can tilt answers toward the demographics and norms of certain communities, potentially drawing out expert and underrepresented views. You can imagine that in areas like medicine, law, or science, this content can sometimes be misaligned, which can cause confusion and misinformation.
The way I see this is going is that LLMs are changing from “answer engines” to “evidence navigators”, because answers will have citations and maybe confidence bands, especially on contested topics. It’s crucial that the next phase of LLMs will have to focus on a type of controlled retrieval, maybe licensing data, and richer provenance so that the models can be both up-to-date and trustworthy.
The truth is, if you want world-class talent, you can’t just hope they’ll come to you. You have to go out and earn their attention—and that takes a completely different approach. So, how do you actually find and hire world-class talent?
Will AI replace Human Thinking? — I think the case for writing and coding manually is still legitimate. You can vibe code everything, and you shouldn’t. [LINK]
Atlassian is acquiring The Browser Company of New York — as an avid user of their browsers (Arc + Dia), I’m stoked to see this finally happen. This is a bold move by Atlassian, but I think pretty clever one as they’re trying to build an all-in-one browser app for the knowledge worker. [LINK]
Flight Connections app — probably one of the best UX/UI interfaces I’ve seen when it comes to looking for flights. It’s also nice to see where you could go if you were feeling lucky or spontaneous. [LINK]
Google is developing its own Layer 1 blockchain called GCUL — this is the beginning of the blockchain wars on payments. [LINK]
Google’s comeback — really interesting to see how the company has climbed its way back over the past 3-4 years. Don’t bet against Google. [X]
Honking cultures — interesting to see Namibia and Botswana among the few countries in Africa that don’t view honking as common behavior. [LINK]
The physics of bamdinton’s new killer spin serve — but it’s been banned because it’s nearly impossible to return. [LINK, VIDEO]
📡 The Signal
Stripe is building their own blockchain: Tempo
Stripe is going hard on crypto. I’ve mentioned Stripe a few times across a number of TWS newsletters, and now the momentum is picking up. Basically, Stripe and Paradigm are launching Tempo, which is an Ethereum-compatible Layer 1 focused purely on payments—you should check out this nice Investopedia piece to explain what the Layer 1 stuff is about. The goal is to support 100k+ transactions per second, a built-in stablecoin, optional privacy, and even “stablecoin neutrality” so that any stablecoin can be used for gas and payments. Since Stripe already processes $1.4T a year in payments, they are now moving into stablecoins to optimize cross-border payments. Crypto is back, but it’s more legitimized, showing proper real-world use cases that are finally coming to fruition for the finance industry. I think this will be one of the chess moves that will finally make on-chain payments normal — paying for goods/services (like your coffee) with crypto. [LINK]
Google releases Gemini-2.5 Flash Image (aka nano-banana)
AI image generation has come a really long way, but it still has its quirks, like keeping things consistent. You want your character or product to look the same, but usually, the results are all over the place. Google’s latest attempt with nano-banana lets you blend multiple images together, edit stuff just by telling it what you want in plain language, and the big one: keeps your images looking the same no matter how many times you change the scene or angle. The key innovation is that it understands what’s in the image thanks to its “world knowledge”. That means it can read hand-drawn diagrams, answer questions about real-world objects, and follow complicated instructions without getting confused.
Some will say this is pretty gimmicky and has no real-world value. I’d beg to differ — this can translate well to video, since it’s just a series of images stitched together, played at 24-30 FPS. This will allow for characters to have the same likeness across multiple scenes. [LINK]
China Releases “AI Plus” Policy
China is mandating a hard adoption of AI. The government wants over 70% of people using AI assistants/agents/devices by 2027, and 90% by 2030. It lays out six action tracks across science, industry, consumer, public services, governance, and global under eight support pillars (models, data, compute, apps, open source, talent, policy, security). The ultimate goal here is to treat AI as basic infrastructure and make it normal in cities, schools, clinics, and even government processes. You should read more about their Internet Plus policy, which builds on top of this. By encouraging people to get more familiar with AI, the government is hoping that the younger generation, especially, will be building the country’s future with the help of AI, which leads to building at faster speeds, more efficiency, and ultimately more prosperity. [LINK]
The rise of White-collar PEDs
Yep, this is a real thing, and it’s getting scarily popular from Wall Street all the way to Silicon Valley. It’s hard to ignore that the current narrative is to hustle, maximize productivity, and be locked in. Don’t get me wrong, I’m all for this, but it’s concerning to see how many people are willing to take PEDs to enhance their cognitive durability and capacity. The author (Park MacDougald) sets the stage here and becomes the guinea pig by subjecting himself to protocols of both pills and powders (Vynanse, Adderall cousins, modafinil, nicotine) and also the wellness stack (sunlight, cold showers, strict sleep, no alcohol, and reduced sugar). The TLDR on this is that Vynanse works well for focus, but the hangover and insomnia are pretty brutal. Strattera kind of works without the dopamine hit, and feels like a joyless slog with a surprise mood crash. Dextroamphetamine is super buzzy and social, but erratic and spiky. Adrafinil/modafinil gives clean wakefulness and fewer crashes, but tolerance and headaches show up fast. On the other hand, Park reports that the non‑drug protocol (early sunlight, daily training, sauna + cold, tighter screen limits, earlier sleep, no booze) actually makes him feel great. However, it’s fair to say that there is no wonder drug that does it all. Even the more natural pathways are not permanent. It takes discipline, consistency, and self-awareness, and knowing that you can’t be productive 24/7. [LINK]
Building India’s AI backbone
In TWS:#021, I mentioned how India is playing catch-up to build their own national AI model. Now, Reliance Industries (owned by Mukesh Ambani — India’s richest man) is trying a different path: building the entire stack end-to-end locally. The plan is for them to launch “Reliance Intelligence” to create a dedicated AI cloud region with Google, then pair with Meta to sell a Llama-powered platform-as-a-service for real enterprise uses like sales, support, and finance. This might actually work since Reliance controls distribution (via Jio — its telco provider), has energy to power data centers, and can bundle pricing across telco, cloud, and its apps. I feel like this is more than “AI infrastructure”, but it’s a carrier-grade go-to-market for Reliance, since they can preload AI into networks, devices, and media, then upsell into enterprises. The capital is there, but the question will be on AI talent and seeing if they can build a world-class team of engineers and researchers to build out local AI models for the Indian market. [LINK]
Australia is making a lunar rover
It’s always nice when you see Australia represented on the global stage, especially when it comes to technology. The ELO2 Roo-ver is a joint project between NASA and the Australian Space Agency (ASA) to co-build a rover for the Moon by 2020. The plan is to collect lunar soil and run experiments to extract oxygen from the atmosphere (if any). Australia is leading in its strength in remote ops, mining tech, and harsh-environment robotics, which is exactly what the lunar surface demands. The second-order effect here is to see spinoffs in advanced manufacturing, sensors, and AI innovations take place. [LINK, ASA]
Australia’s big social media ban is coming — but do we really need it?
On the other side of the spectrum, Australia is planning on rolling out its policy to ban social media for under-16s by December this year and force platforms to verify ages. The problem with this is going to be enforcement. I’m personally not a big fan of this policy because I think the government shouldn’t be intervening in matters where this should be taken care of inside the familial unit. Nonetheless, there will no doubt be loopholes to this, where teens will be able to find backdoors. The main point here is that I don’t think anyone, i.e., parents, kids, or even the platforms, has the clarity on what verification will actually be used, who holds the data, or how mistakes get fixed. [LINK]
If you liked this, here are a few more popular posts:
