Codex can be used to file personal taxes, and it can even be more accurate than a human accountant. The immediate feedback that Codex gives users helps them understand their situation and the tax code better. Give Codex access to personal details, and it can make better assumptions about tax treatment. Accountants probably don't have to worry - most don't enjoy doing personal tax returns, and their other duties are likely much harder to automate.
Open source models offer the proposition of distributing the value created by AI more broadly, enabling more people to build. However, the current ecosystem doesn't make building easy. The stack contains many bugs, and there's debt hidden in every layer. Developers need to stop patching bugs and build a better stack.
ChatGPT's engagement lead is wider than its market share lead: DAU:MAU sits at 45% versus Gemini's 22%, and WAU:MAU has climbed from 50% in mid-2023 to 82% today, putting it ahead of Gmail and Spotify and approaching Instagram. Week 4 retention of 66% beats every enterprise app in the dataset and is more than double Perplexity's 24%, while most apps see retention flatten at scale. The rarest signal is ChatGPT's smile curve: one of only three products alongside Gmail and Chrome where retention dips and then recovers, meaning OpenAI's product releases are actively reactivating lapsed users, not just attracting new ones.
Benchmarks are trotted out whenever a new model is released, but what do they actually measure? ngrok's Sam Rose dug into the papers, code, and critiques to understand what 14 popular AI benchmarks actually mean [12]. E.g., SWE-bench Verified = fixing small bugs in 12 popular
Agentic coding benchmarks are commonly used to compare the software engineering capabilities of frontier models. These scores are often treated as precise measurements of relative model capability. However, research shows that infrastructure configuration alone can produce significant differences in scores. Eval developers have begun accounting for this, but current fixes potentially change what benchmarks end up actually measuring.
RCLI is an on-device voice AI for macOS that can control apps and perform other actions via voice. Users can choose from a variety of local AIs and perform 38 macOS actions. The tool can be used to ingest local documents for users to query. RCLI is open source under the MIT license.
An agent is a model with a harness. Harness engineering is the process of turning models into work engines by building systems around them. Models contain the intelligence, and the harness makes that intelligence useful. This post looks at what harnesses are and details the core opponents that agents need.
Every AI training pipeline rests on a data layer that determines how those models behave. This data determines what the models know, how they reason, and what they can safely do. However, much of today's training data remains opaque, fragmented, or siloed across teams. Open data access gives developers a faster and more cost-effective path to building high-quality models. This is why Nvidia releases open datasets along with its open models, tools, and training techniques.
Perplexity's Comet AI browser has been blocked from accessing Amazon's site. Amazon sued Perplexity in November, alleging the startup concealed its AI agents so it could continue to scrape Amazon's website without approval. The Comet browser allows users to ask the assistant to find items on Amazon and make purchases. Amazon says Perplexity's agents pose a risk to customer data and that they create challenges for its advertising business. It has broadly locked down its shopping sites from AI agents and blocked dozens of agents.
Frontier labs are continuing to build first-party products on top of their own capabilities. For every token they sell, they can spend it themselves for COGs of under 50%. Your supplier is also your competitor, and they have a permanent cost advantage over you. Applied AI companies need to win on something other than tokens. Businesses need to build products so good that the token itself becomes the commodity input, not the value proposition.
OpenAI's IH‑Challenge is a dataset designed to train models to prioritize instructions based on trust level across system prompts, developers, users, and external data.
claude-ground introduces a minimal rule system for Claude Code, providing phase tracking, decision logging, and language-specific best practices to improve coding discipline.
CFTC Chair Michael Selig announced the Project Crypto Initiative with the SEC to establish a clear asset taxonomy. The plan includes registration guidance for DeFi developers and rules for prediction markets. This shift from enforcement-led oversight aims to solidify the US as a global digital asset hub.
Circle's Nanopayments infrastructure enables gas-free, sub-cent USDC transactions for AI agents using the x402 open standard. By aggregating off-chain payments for periodic on-chain settlement, the system supports machine-to-machine economies across networks like Arbitrum and Base, addressing cost barriers for high-frequency, micro-scale digital services.
Virtuals Protocol and the Ethereum Foundation's dAI team have co-developed ERC-8183, an open standard for trustless AI agent commerce that encodes escrow and evaluator attestation as onchain smart contracts. The standard's core primitive is the "Job," a three-party construct (Client, Provider, and Evaluator) progressing through a defined sequence – Open, Funded, Submitted, then Terminal – where the Evaluator role can be filled by an AI agent for subjective tasks, a ZK verifier for deterministic outputs, or a multi-sig for high-stakes agreements. Composability is achieved through optional "Hook" contracts enabling custom logic such as reputation-gating, private job posting, and dynamic bidding, while the standard integrates symbiotically with ERC-8004 to form a full agent economy stack: identity and discovery (ERC-8004) feed into commerce (ERC-8183), which feeds back into reputation (ERC-8004). ERC-8183 is explicitly positioned as complementary to x402 rather than competitive – x402 handles HTTP-layer micropayments while ERC-8183 governs the full commerce lifecycle, including job specification, escrow, deliverable submission, attestation, and settlement.
Wildcat, an undercollateralised credit protocol on Ethereum that has facilitated several hundred million dollars in onchain credit, has announced V2.5 alongside three new product lines. V2.5 transitions the core underwriting model from deposit pools to revolving credit facilities (RCFs) with APR split into commitment and utilisation components, and introduces sealed-bid auctions for borrower price discovery alongside precommitment levels and rolling lockup hooks. New product lines include WildUSD/CatUSD, a tranched CLO system with senior/junior tranches designed to produce a credit-backed stablecoin, ERC-4626 wrappers for non-rebasing debt tokens, cross-asset collateral vaults, and both covered and naked credit default swaps.
The stablecoin market has reached $266 billion in supply and processed $374 billion in real payments last year – a 76% year-over-year increase that, notably, decoupled from broader crypto market cycles for the first time. USDT and USDC still command 97% of supply, leaving remaining entrants competing on an increasingly commoditized axis of rails, speed, compliance, and feature sets. Underlying payment data demonstrates a shift toward everyday commerce: average transaction size has compressed from $402 to $342 while consumer-to-business volumes grew 131%, pulling in users with no native crypto orientation. The differentiation question that remains unresolved is whether the next dominant stablecoin will be won on brand and meaning rather than basis points, and whether AI agents making autonomous financial decisions can be influenced by non-cost factors in stablecoin selection.
Crypto adoption will shift from “buying in” to “earning in,” as stablecoins and crypto-native rewards become a primary income rail for globally distributed contributors and turn earnings into balances, savings, and eventually participation in onchain finance. They define Internet Labor Markets as contributor-owned marketplaces where verifiable, dynamically created tasks settle instantly over crypto rails, enabled by bespoke task inventory plus deterministic and bonded verification to keep payouts aligned with real work. AI is framed as the catalyst that shrinks core teams and decomposes jobs into modular tasks, making sourcing, verifying, and paying for marginal contributions the new bottleneck that ILMs are built to solve.
Token launch performance data from 2023-2025 reveals systemic failure, as all 28 tokens launched at $1B+ FDV traded below listing price with median drawdown of -81%, while 85% of all tokens ended 2025 negative at -69% median. The root cause is structural – airdrop mechanics distribute tokens at zero cost basis, compounded by Sybil exploitation. Only 9.4% of tokens declining in their first week recovered by year-end, suggesting early distribution quality is deterministic for long-term outcomes. The industry must shift from open distribution to deliberate cap table construction using onchain history as filtering infrastructure.
Tokenized gold products PAXG and XAUT, plus Hyperliquid's HIP-3 perpetuals, facilitated 24/7 price discovery during geopolitical shocks. With $6.1B in tokenized gold and $95B in perp volume, these rails enable continuous macro hedging for commodities and equities when traditional exchanges are closed.
Anthropic's newly released 33-page Claude skills documentation includes a prediction market trading bot template claiming a 68.4% win rate and $300-$1,500 daily returns.
Nunchi released a standalone CLI trading agent for Hyperliquid featuring 14 purpose-built strategies for yield perps on YEX, compatible with Claude, OpenAI, and OpenClaw frameworks.
A federal judge in Ohio denied Kalshi's bid to stop state enforcement, saying the record shows Congress did not intend the Commodity Exchange Act to preempt state sports gambling laws, clearing the way for Ohio to treat Kalshi's sports event contracts as illegal sports betting.
Apple is promoting the MacBook Neo to first-time Mac users with a quirky TikTok campaign aimed at Gen Alpha. The campaign features surreal “brain rot” style videos and introduces a tiny character based on the macOS Finder icon, briefly appearing at a miniature desk with a tiny MacBook Neo. The character has quickly caught the attention of Apple fans, who have already started creating their own versions online.
GitHub Copilot users can now connect to Figma's MCP server to push rendered UI to the canvas as editable frames and pull design context into code. The integration is now available in VS Code and will be coming soon to the Copilot CLI. This feature helps users scope new features more easily, explore design directions, and generate code that matches existing designs.
The Pixel 11 Pro Fold leak suggests only a small design update, with slightly reduced thickness but still thicker than competing foldables. Rivals are focusing heavily on thinner hardware, so Google will need to stand out through stronger performance, battery life, cameras, AI features, and superior foldable software. Without clear advantages in those areas, the device risks feeling like an incremental update in a rapidly evolving market.
Waymo, despite its self-driving technology, pays DoorDash drivers to close car doors manually left open by passengers instead of implementing automated door systems that already exist. This reflects a broader problem in tech: companies choose inefficient solutions and overlook obvious design alternatives when rushing products to market and prioritizing cost-cutting over user experience. The design industry has shifted from solving problems efficiently to simply following corporate directives that prioritize profits over reducing customer friction.
Experience strategists work at the executive level, focusing on organizational priorities rather than just designing products like other UX roles. They identify how improving customer and employee experiences can solve executive challenges, then orchestrate solutions that achieve strategic business goals. For example, one strategist studied small business owners' payment delays and connected this insight to executives' revenue concerns, creating strategic opportunities that elevated UX beyond just "making things pretty."
Automated accessibility scanners can quickly identify basic issues on webpages but typically detect only about 40% of problems and cannot evaluate real usability, making them insufficient on their own. Effective accessibility evaluation requires combining automated tools with manual checks (for structure and semantics) and experiential testing tools (that simulate disabilities) to ensure users can truly understand, navigate, and interact with a site.
Sprite Studio is an AI-powered pixel art generator that creates sprite sheet animations for games based on character or object descriptions. Users can adjust grid layouts, frame rates, and preview their animations with various canvas backgrounds before exporting the final sprite sheets.
HeyVid AI is a free online platform that provides access to top AI video and image generation models, including Kling AI, Sora 2, Runway, Midjourney, and Flux AI.
Maizzle is a framework that helps you quickly build HTML emails with Tailwind CSS, including advanced email optimizations like CSS inlining. It allows you to use familiar HTML markup and utility-first CSS classes to style emails faster while maintaining control over responsive design.
AI-generated UX interfaces look polished at first, but feel wrong in actual use because AI lacks the contextual judgment and human intuition that professional designers rely on to create truly usable experiences. While AI excels at pattern matching and visual hierarchy rules, it cannot anticipate user hesitation, weigh competing priorities, or make design decisions based on real user behavior insights. The speed of AI generation actually creates problems, since good UX requires deliberate slowness for testing, revision, and consideration of constraints that make design options appropriate for specific contexts.
Complexity often sells better than simplicity, despite simplicity being a virtue that requires hard work to achieve. In many engineering environments, employees who over-engineer solutions get promoted because their complex work creates compelling narratives. This incentive problem starts in interviews where complex system designs impress more than straightforward solutions, and continues through design reviews where engineers feel pressured to add unnecessary abstractions and future-proofing.
Hippeas has refreshed its packaging to evolve from a minimalist “iconic” design into a more functional retail brand, introducing a chickpea mascot, a broader color system, and prominent product photography to improve flavor differentiation and highlight taste. The redesign prioritizes shelf clarity, stronger product appeal, and clearer messaging while still preserving the brand's signature yellow identity as it scales across more products.
Persuasive design has matured over the past decade into "behavioral design," moving beyond gamification tricks toward frameworks like COM-B that account for capability, opportunity, and motivation in context.
Apple's iOS 26.4 beta adds eight new emojis—including a “distorted face” that resembles the character from Apple's controversial “Crush” ad—sparking mixed reactions from fans who see it either as a humorous self-reference or an unnecessary reminder of the criticized campaign.
Each level of agentic engineering (using agents effectively) represents a leap in output, with a "multiplayer effect" encouraging teams to advance together to maximize efficiency. The progression starts with basic AI assistance like tab-completion and AI-focused IDEs, then moves into more sophisticated practices such as context engineering, compounding engineering, and integrating MCPs and custom skills. The highest levels involve building autonomous agents with automated feedback loops.
On February 17, an attacker published a malicious cline@2.3.0 to npm using a stolen publish token, injecting a postinstall hook that silently installed OpenClaw (a background AI daemon with full disk/terminal access) on ~4,000 machines over 8 hours. The theft was pulled off via a chain of prompt injection through Cline's AI issue triage bot. The fix had already been published by a researcher 8 days earlier, but Cline revoked the wrong token, leaving the npm publish token alive for a second actor to exploit.
Micro frontends are really an org problem disguised as a tech problem — the actual win is letting teams own their vertical slice end-to-end without coordinating deployments. However, the complexity cost (separate CI, dependency versioning hell, and consistency overhead) isn't worth it unless you have 5+ teams actively stepping on each other. For most teams, a well-structured monorepo gets you 90% of the benefit with none of the pain.
This dev built his own customized text editor over two years. He dogfooded it himself, using his editor daily despite its early pains, which drove rapid development and feature implementation over recent months. He found a lot of satisfaction and increased long-term productivity from building a tool perfectly tailored to his needs.
All the optional best practices that devs commonly skip, such as 100% test coverage, small well-scoped files, end-to-end types, and fast ephemeral dev environments, turn out to be non-optional when agents are coding. Agents can't navigate a messy codebase and amplify messiness.