Amazon is now the biggest global company by revenue. Walmart was previously the leader for more than a decade. Jeff Bezos had carefully studied Walmart's founder, Sam Walton, and embedded many of his business strategies while building Amazon. Amazon's revenue increased at almost 10 times the pace of Walmart's over the past decade. Amazon is the biggest online retailer, attracting 2.7 billion visits to its websites and mobile apps each month, while Walmart is the biggest physical retailer in the world, with more than 10,000 stores and shopping clubs globally.
2026.02.20·links.tldrnewsletter.com·3 min read·TLDR Scaling laws have defined tech in the last five years. Large language models have demonstrated consistent returns to scaling. However, these 'laws' apply to very few other fields. Little work has been done to systematically understand, in theory or in general, what causes scaling laws or determines their slope and robustness. This article looks at what causes scaling laws, what their preconditions are, when they stop, why they have worked in some fields but not others, and more.
2026.02.20·mackenziemorehead.com·24 min read·TLDR Almost everyone agrees that AI policy should be evidence-based. However, science is often messy and full of caveats, which doesn't fit well with policymakers' demands for clean answers. Many AI benchmarks have saturated and can no longer give a clear 'no' on risks. Experiments show how much we still have to learn. We need to spend less time proving that today's AIs are safe and more time figuring out how to tell if future AIs are dangerous.
Sponsored
Traffic skyrocketing? If you're on Microsoft Azure [10], there's no reason to panic. When user demand spikes, Azure adjusts capacity in the background. Performance stays steady with no manual intervention, no late-night firefighting, and no scrambling to explain why the site
2026.02.20·links.tldrnewsletter.com·Sponsor·TLDR 2026.02.20·links.tldrnewsletter.com·TLDR The winning formula in early 2026 is structured context, tiered rigor, smaller teams with higher leverage, and relentless measurement of downstream effects. The teams that adapt AI tools without the underlying taste and discipline are discovering that AI just makes their existing problems bigger. Teams with strong engineering cultures, robust CI/CD, clear architectural standards, and effective review processes see AI compound their advantages. This is widening the gap between elite and average teams.
2026.02.20·cjroth.com·20 min read·TLDR AI should be treated as an amplifier of human capacity rather than a replacement. AI can still act autonomously with specific tasks, but in a way that is an extension of human decision-making and context. This allows humans to do dramatically more work, more sustainably, with less injury and fatigue. The AI industry's dream of having a fully autonomous AI employee is seductive, but it will likely set us up for disappointment.
2026.02.20·kasava.dev·14 min read·TLDR OpenAI's technology isn't unique. While it has a big userbase, there is limited engagement and stickiness, and no network effects. Its competitors have matched its technology and are leveraging their product distribution. Executing better than everyone else is certainly an aspiration, but it isn't a strategy.
2026.02.20·ben-evans.com·24 min read·TLDR ByteDance is hiring nearly 100 employees for its AI division in the US. The company's AI team, called Seed, was established in 2023 and has labs across the US, Singapore, and China. The open roles include various job responsibilities such as producing international data, advancing ByteDance's text, image, and video generation tools, AI research, and building models for drug discovery and design. The company's recent AI launches have thrust it into the spotlight in the US. There has been some controversy over the data used to train its models.
2026.02.20·links.tldrnewsletter.com·5 min read·TLDR Sponsored
Tech stack complexity is out of control, and the last thing you need is a bigger ‘stack'. Microsoft Azure gives you a unified view of applications, cloud resources, and AI—so you can ship more software
2026.02.20·links.tldrnewsletter.com·Sponsor·TLDR 2026.02.20·links.tldrnewsletter.com·TLDR Meta is separating its Quest VR platform from its Worlds platform, shifting the focus of Worlds to be almost exclusively mobile.
2026.02.20·engadget.com·1 min read·TLDR There were around 1,500 federal trade secrets cases filed in the US last year, up 20% from the previous year.
2026.02.20·links.tldrnewsletter.com·4 min read·TLDR Software that can be generated in a weekend is disposable inventory, not something you can build a startup around.
The productivity gains from AI are real, but it won't be catastrophic, as it will just devalue mechanical skills while elevating judgment skills.
2026.02.20·weightythoughts.com·20 min read·TLDR Nine observations from building with AI agents over the last year.
2026.02.20·tomtunguz.com·3 min read·TLDR The most value has always historically accrued to the companies that did the hardest things.
2026.02.20·danhock.co·13 min read·TLDR Sponsored
Pliny the Liberator is AI's most notorious jailbreak researcher, known for turning “harmless” LLMs into offensive tools.
Google released Gemini 3.1 Pro as the upgraded core model behind recent Gemini 3 “Deep Think” improvements, and began rolling it out to the Gemini API/AI Studio, Vertex AI, Android Studio, the Gemini app, and NotebookLM. The post highlighted a verified 77.1% score on ARC-AGI-2, more than doubling Gemini 3 Pro's result.
2026.02.20·blog.google·5 min read·TLDR AI Google tests NotebookLM integration within Opal workflows, enhancing data extraction and automation. This integration aims to streamline processes and improve workflow efficiency for users.
2026.02.20·testingcatalog.com·1 min read·TLDR AI ARC-AGI-3 is an Interactive Reasoning Benchmark designed to measure an AI Agent's ability to generalize in novel, unseen environments. Opus 4.6 demonstrates better reasoning and use of memory than Gemini 3.1 Pro and solves more levels. Current models may be able to solve ARC-AGI-3 given access to a harness with a simple memory. Memory scaffolds are likely enough for pseudo-continual learning to push us to some self-improvement or research-agent threshold within the next 2 years.
2026.02.20·links.tldrnewsletter.com·5 min read·TLDR AI 2026.02.20·thezvi.wordpress.com·58 min read·TLDR AI Cursor described an “agent sandboxing” system that let local coding agents run freely inside a constrained environment and only asked for approval when leaving the sandbox (often for internet access).
2026.02.20·cursor.com·7 min read·TLDR AI Sponsored
Looking to deploy AI that you actually own? Crusoe Managed Inference [12] unlocks breakthrough speed and throughput without the infra overhead. Run SOTA fine-tuned models: DeepSeek, gpt-oss, Kimi, or bring your own. Power production apps with high reliability. Focus on innovation and leave the clusters to Crusoe's expert team. Try a model
2026.02.20·crusoe.ai·Sponsor·TLDR AI optimize_anything is a declarative API that optimizes any artifact representable as text. Users declare what to optimize and how to measure it, and the system handles the search. It consistently matches or outperforms domain-specific tools. A surprisingly wide range of problems can be formulated as optimizing a text artifact. If it can be serialized to a string and its quality measured, a large language model can reason about it and propose improvements.
2026.02.20·gepa-ai.github.io·132 min read·TLDR AI When not using reasoning, repeating the input prompt improves performance for popular models without increasing the number of generated tokens or latency. It is interesting that tricks like this are still possible despite the amount of work being put into improving large language models. The discovery proves how much room for improvement there still is for current models.
2026.02.20·daoudclarke.net·1 min read·TLDR AI Building on Google's Transformer architecture, the authors proposed training sequence-model agents against many different opponents so they learned to adapt within each game without hardcoded assumptions about how others learn.
2026.02.20·arxiv.org·9 min read·TLDR AI OpenAI has a big user base, but it has limited engagement and stickiness and no network effect. The company doesn't have any unique technology. The incumbents have already matched the technology and are leveraging their product and distribution. This post takes a look at OpenAI's strategy and how the company can compete in today's landscape.
2026.02.20·ben-evans.com·25 min read·TLDR AI Prototype with the best and polish small gems. Use teams of agents as micromanagers, and experiment with different tools and workflows. Document everything to create improvement loops that improve success rates without manual intervention. Skills are easier to debug than code.
2026.02.20·tomtunguz.com·2 min read·TLDR AI Sam Altman and Dario Amodei avoided hand-holding at an AI summit, highlighting their tense rivalry.
2026.02.20·fortune.com·5 min read·TLDR AI DuckDuckGo has launched AI-powered image editing on Duck.ai, allowing users to edit images without needing an account, though subscribers have higher daily limits.
2026.02.20·9to5mac.com·2 min read·TLDR AI Apex Group, which oversees $3.5 trillion in assets, announced a partnership with Trump-affiliated World Liberty Financial to pilot the USD1 stablecoin for fund subscriptions, redemptions, and distributions across its tokenized fund ecosystem. The collaboration will also explore making WLFI's tokenized real estate and infrastructure assets available on the London Stock Exchange Group's Digital Market Infrastructure platform, pending regulatory approval. Apex has been expanding its tokenization capabilities through recent acquisitions of Tokeny and Globacap, and is well-positioned to serve institutional clients seeking blockchain-based settlement efficiency and cost reduction.
Uniswap governance is voting on a temp check to activate protocol fees across all remaining v3 pools and roll fees out to eight additional chains (Arbitrum, Base, Celo, OP Mainnet, Soneium, X Layer, Worldchain, and Zora). They are also voting to introduce a tier-based adapter to auto-apply fees by LP fee tier, and route collected revenue through chain-specific TokenJars back to Ethereum for conversion and UNI burns under the post-UNIfication fee-and-burn framework.
2026.02.20·links.tldrnewsletter.com·2 min read·TLDR Crypto There is a five-step roadmap to make Ethereum the first LLM-driven blockchain, spanning the full governance stack from validator delegation of upgrade decisions to AI agents, through LLM-assisted EIP authoring and review, to ultimately AI-generated client codebases that are formally verified and fully test-covered. Ethereum holds a structural advantage given that its EIP corpus, ACD call transcripts, and open governance discussions constitute high-quality training data already embedded in frontier models. The Ethereum Foundation has signaled institutional alignment by hiring tooling coordinators and forming a dedicated dAI team.
PrivateX402 introduces privacy-preserving payment channels for multi-agent AI systems, concealing individual payment details from on-chain observers while enabling users to allocate budgets across multiple agents. The protocol requires only two onchain transactions – setup and settlement – with all intermediate payments occurring off-chain at no additional cost. By replacing centralized payment facilitators with a verifiable computation layer and integrated dispute resolution, PrivateX402 makes high-frequency AI agent micropayments economically viable.
Vitalik Buterin drew a sharp distinction between Ethereum as a permissionless, censorship-resistant protocol and the separate obligation of people and communities to hold and articulate clear principles. Protocol-level neutrality does not extend to individuals, and criticism within the ecosystem is not censorship. Using Linux as an analogy, open neutral infrastructure can still serve corporate ends, meaning those who care about user empowerment must proactively build full-stack ecosystems consistent with those values rather than assuming protocol neutrality alone is sufficient.
x402's payment rails function correctly, but wallet onboarding is broken across three dominant patterns – plaintext local keys with no recovery, per-app embedded wallets that fragment funds, and managed wallets that lock users inside a single provider's infrastructure. The proposed fix inverts the current model: rather than applications provisioning wallets, users should bring a single portable wallet to every x402-compatible service, with one identity and one key set universally accessible. Para is a reference implementation of this architecture, where users own keys, email recovery is native, and the wallet is fully decoupled from the application layer.
2026.02.20·links.tldrnewsletter.com·4 min read·TLDR Crypto 2026.02.20·links.tldrnewsletter.com·7 min read·TLDR Crypto Aave founder Stani Kulechov outlined a thesis positioning Aave as the primary DeFi financing layer for solar energy infrastructure, arguing that traditional capital markets are structurally overexposed to scarcity-based assets while a $30-50 trillion investment gap in abundance technologies remains largely unaddressed through 2050. The proposed mechanism would have Aave accept tokenized solar project debt as collateral, enabling developers to borrow stablecoins against receivables within minutes – a $100M tokenized portfolio could unlock $70M in GHO instantly. Solar's financial profile makes it particularly well-suited, as projects already achieve 70-85% debt financing with unlevered returns around 8% that lever up to 15-20%.
2026.02.20·links.tldrnewsletter.com·8 min read·TLDR Crypto AI agents require native payment capabilities for true autonomy.
2026.02.20·links.tldrnewsletter.com·6 min read·TLDR Crypto BlackRock integrated its BUIDL fund into UniswapX via Securitize, while Apollo Global Management partnered with Morpho for onchain lending.
Base is moving to a unified stack built on Reth to accelerate development.
Robinhood's Arbitrum-based Layer-2 testnet logged four million transactions in its first week.
2026.02.20·links.tldrnewsletter.com·1 min read·TLDR Crypto Canva closed 2025 with $4 billion in annual recurring revenue, over 265 million monthly active users, and a 20% increase in MAUs driven partly by AI tool adoption. The company is pivoting from a design platform with AI features to an "AI platform with design tools," while integrating with chatbots like ChatGPT and Claude. LLM referral traffic is now growing in double-digit percentages, and Canva is actively investing in surfacing its content through AI-powered search.
Variant's new AI eyedropper feature lets users click on one generated interface to transfer its style onto another. However, the tool's usefulness is limited because Variant's AI engine produces designs that already look too similar, and it can't sample from external sources like bitmaps or Figma. The feature nonetheless points toward a promising direction: giving designers direct manipulation tools for AI instead of relying on imprecise text prompts.
2026.02.20·links.tldrnewsletter.com·4 min read·TLDR Design Ahead of Apple's 4 March event, Nothing cheekily hijacked the hype by spray-painting “Nothing” and “5th March” over Apple's invite, likely teasing the Nothing Phone (4a) Pro launch a day later. While some fans praised the bold, counter-culture stunt, others slammed it as unprofessional, though the move reinforces Nothing's provocative brand identity and willingness to challenge Apple's dominance.
RITE (Rapid Iterative Testing and Evaluation) offers a scrappier, cheaper alternative to formalized UX research by testing with smaller groups and iterating prototypes in real time. Traditional user testing has become too expensive and siloed, often producing reports rather than actionable solutions, making it vulnerable to AI-based alternatives. By focusing on co-creating use cases with real users early in the process, RITE testing helps teams build products people actually want to buy.
2026.02.20·dataanddesign.substack.com·9 min read·TLDR Design Apple has treated 120Hz ProMotion displays as a premium feature for nearly a decade, limiting them mostly to Pro devices due to the smoother experience they provide. Given the large price gaps between mid-range and Pro models, and the fact that Apple's products aren't cheap to begin with, it feels outdated for devices like the iPad Air and MacBook Air to still ship with 60Hz displays instead of at least 90Hz or 120Hz.
Modern personalized systems don't just reduce friction—they increasingly shape what we think about next, optimising for engagement over understanding and quietly eroding “cognitive sovereignty,” the ability to explore and decide independently. The challenge isn't personalization itself, but business models and design defaults that prioritize retention over agency, making it crucial to build systems with transparency, user control, friction where needed, and space for genuine discovery rather than tightly steered attention.
2026.02.20·links.tldrnewsletter.com·11 min read·TLDR Design