News Feed

The article explores 'steering'—tuning an LLM’s activations during inference to bias its behavior—through DeepSeek-V4-Flash and the DwarfStar 4 project, outlining simple and more sophisticated ways to identify and boost activation-based concepts like “respond tersely.” It argues that steering could offer a cleaner alternative to prompt engineering but is limited by access to weights, open-model performance, and the fact that many gains can be achieved via prompting or fine-tuning, leaving practical adoption uncertain. It closes by noting the open-source community may eventually build a library of boostable features if interest grows.

Eric Jang discusses rebuilding AlphaGo from scratch, outlining why Go's deep, large game tree made brute-force search impractical and how AlphaGo uses neural nets to prune and evaluate moves within a Monte Carlo Tree Search framework (PUCT, Q(s,a), visit counts). He also explains Go rule variants, scoring differences between human and computer play, and notes that modern open-source progress (e.g., KataGo) and tools like LLMs have dramatically reduced the compute needed to train strong Go AIs, enabling more accessible research.

Memory has become the new bottleneck for AI infrastructure, with DRAM and NAND prices surging since last year and DRAM up more than 2x year-over-year as of March, boosting profits for memory makers (Samsung, SK Hynix, Micron). With hyperscalers locking in multi-year supply agreements and capacity expansions lagging, memory could sustain higher pricing power and drive a 10–20% price lift on consumer devices like PCs and phones. The piece also notes AI-driven productivity gains (e.g., Affirm’s AI-first retooling) and discusses broader implications for AI adoption, labor, and markets, highlighting memory as a strategic bottleneck rather than a pure commodity.

Weekly Dose of Optimism #193 highlights major biotech and space-tech news, including Isomorphic Labs' $2.1B Series B to advance AI-driven drug design and a space-drug manufacturing collaboration between Varda and United Therapeutics. It also covers Revolution Medicines' daraxonrasib showing promising results against pancreatic cancer and a wave of IPO activity (Cerebras, Fervo) alongside the quirky space-tech rebrand Cowboy Space Corp. Overall, the issue frames AI-enabled drug discovery, space-based manufacturing, and renewed venture liquidity as converging forces driving optimism about the next frontier of biotech and tech.

The article covers closing arguments in the Musk v. Altman trial, highlighting missteps by Musk’s lawyer and OpenAI’s methodical presentation of evidence, while interspersing gossip-driven color. It then offers substantive takeaways, arguing that Musk’s AI efforts (notably xAI) are financially draining and underperforming, with internal tensions and credibility questions among top players. Overall, it blends courtroom narrative with analysis of the AI ecosystem and the trial’s implications for OpenAI, Altman, Murati, and Musk.

A new independent institute, the Youth AI Safety Institute, backed by Margrethe Vestager and hosted by Common Sense Media, will be launched in Denmark to provide 'crash-test' style safety ratings for AI used by children. While its exact testing methods for evolving chatbots remain unclear, the move comes amid concerns that current AI products fail to recognize mental-health distress and can deliver delayed crisis alerts, underscoring a call for transparent safety standards. Funded by philanthropies and tech industry partners—including Anthropic, the OpenAI Foundation, and Pinterest—the institute plans open-source evaluation tools and to offer its assessments with editorial independence.

The speaker argues that shrinking the Fed's balance sheet is not the right objective because the Fed's footprint includes its balance sheet plus core functions like bank safety, payment system operation, and Treasury fiscal agency, and that reducing reserves could undermine stability and efficiency. He emphasizes that reserves are safe, liquid, and costless to the Fed, and that the ample-reserves regime has facilitated effective monetary policy and stable markets. The talk warns that proposals to reduce reserve demand or loosen liquidity rules could degrade resilience and policy effectiveness rather than meaningfully shrink the Fed's footprint.

The piece argues that programming languages and platform choices are becoming less locked-in, driven by tools and approaches that ease migration and reversibility. It cites a real-world example of a company rewriting iOS and Android apps in React Native, with the option to port back to native later, illustrating how modern development stacks enable more flexible decisions. The central point is that developers increasingly expect to move between languages and platforms without permanent commitment.

WIRED documents the grim mood at Meta as it plans about 10% layoffs (roughly 8,000 people) and enforces new AI-related workflows, pay cuts, and mandatory monitoring software. The reporting ties morale to court losses, forced role changes for top engineers, rising pay gaps, and aggressive AI spending, with pockets of protest including UK union activity and employee pushback over privacy and consent. While a few groups, like TBD Lab, remain excited by frontier AI, the piece portrays a workplace under pressure with a climate of fear, scrutiny, and upheaval across much of the company.

Mitchell Hashimoto argues that programming languages are becoming fungible, reducing traditional lock-in for projects. He cites Bun’s rapid ability to rewrite or port its code across languages (e.g., Zig to Rust) as evidence that language choice can be swapped quickly. The takeaway is that Rust may be expendable, and long-term viability may hinge more on ecosystem and architecture than on any single language.

DS4 (DwarfStar 4) is celebrated for enabling serious local AI inference with a single-model approach, driven by a fast frontier model and an efficient 2/8-bit quantization that fits 96–128GB RAM. The author envisions the space being defined by the fastest open-weight models on high-end Macs or GPU-in-a-box setups, with DS4 variants tuned for coding, legal, and medical tasks loaded as needed. Looking ahead, priorities include quality benchmarks, a coding agent, CI hardware for long-term reliability, broader ports, and distributed inference, while noting that local models can rival online ones in usefulness through features like vector steering.

Sea Limited is rolling out Codex across its developer organization to enable agentic AI-assisted software development that goes beyond autocomplete, using AI to navigate large microservices, propose test-driven implementations, and reduce technical debt. Internal data show strong adoption (87% weekly active users) and a high likelihood to recommend Codex (73% of frequent users), with AI agents increasingly operating within CI/CD pipelines to boost architecture, debugging, and reliability. Sea frames Southeast Asia as a proving ground for AI-native software and is expanding access through a regional Codex Hackathon series with OpenAI to accelerate AI-driven innovation and talent development.

The speech argues that community banks—particularly in the 10th District—need tailored regulation that reflects size, risk, and business models, rather than a one-size-fits-all approach. It critiques CECL and Regulation O as overly burdensome for small banks and describes a shift to material-risk supervision using clarified MRAs and MRIAs, plus a revised community bank leverage ratio. It also calls for proactive collaboration on emerging technologies like AI and digital assets and highlights local innovations and relationship-based lending as essential strengths.

The article argues that the traditional system of record, especially CRMs, is becoming a data input to a new 'system of intelligence' powered by AI agents and orchestration across the CRM, calendars, emails, and other tools. It contends that the real value will shift upward to the orchestration layer above the database, where foundation models and domain logic enable automated context, decision-making, and action, redefining the GTM software stack. This shift expands the total addressable market for enterprise software, clusters AI-native GTM startups around high-frequency workflows, and promises greater salesperson productivity and preserved institutional memory through automated context.

An online joke about Cliff Stoll’s death reveals how AI hallucinations can spread false information. An AI-generated review on Facebook repeated the claim that he died in May 2024, showing how LLMs can scrape and regurgitate rumors as facts. The piece argues that, in the age of AI, the line between humor and truth is increasingly blurred, echoing Catch-22 and 1984.

Stitch is building an API-first operating system for modern financial institutions, starting in the Middle East, to replace legacy core systems and enable AI-driven capabilities across banks, fintechs, and non-financial players. The piece cites strong early traction (over $5 billion transacted in six months; 10x customer growth in 2025; 20x revenue growth) and GCC expansion into Africa and Southeast Asia, driven by Saudi Arabia’s plan to proliferate financial institutions. It argues Stitch could become the foundational system of record that unlocks AI transformation (loan origination, AI-powered support, streamlined workflows), and notes a16z’s Series A investment in Saudi Arabia as validation of the opportunity.

The article explains how the datasette.io site was protected from poorly-behaved crawlers by building a configurable IP-rate-limiting plugin (created with Codex) and applying it to block IPs that hit critical areas too quickly. It shares the production configuration, including the header used to identify clients, the maximum number of keys, exemption patterns for static assets, and a rule that targets specific endpoints (/global-power-plants/* and /legislators/*) with a 60-second window, 60 requests, and a 20-second block. This demonstrates a practical, field-tested approach to defending a site from abusive crawlers by applying targeted rate limits.

Anthropic's May 2026 policy paper argues democracies must maintain a lead in AI compute and governance to prevent a CCP-led AI frontier by 2028. It sketches two futures: if the US tightens export controls, disrupts distillation, and accelerates democratic AI adoption, it can set global norms and engage China on safety; if those actions are delayed, China could catch up or overtake, enabling widespread AI-enabled repression and military applications. The piece emphasizes compute as the bottleneck, the dangers of a neck-and-neck race, and the value of safety dialogue, while noting early indicators like safety evaluation gaps in Chinese labs and the rapid acceleration tied to 'Mythos Preview' as evidence for urgency.

Genkit announces a middleware system for building AI-powered agentic apps, enabling interception and augmentation of generation, model, and tool execution through composable hooks. It introduces pre-built middleware (Retry, Fallback, ToolApproval, Skills, Filesystem) and shows how to compose and extend behavior with custom middleware like a ContentFilter, including a practical example and a UI for testing and debugging. The article also notes multi-language support (TypeScript, Go, Dart, Python coming soon) and the ability to stack middleware and publish reusable packages.

The interview attributes the recent S&P rally to mechanical, systematic flows from trend-following CTAs and passive investing, including large-scale rebalancing and 401(k) inflows, which produced a record one-month surge and a sharp VIX decline rather than changes in fundamentals. Green cautions that as passive ownership grows, markets may behave like low‑float stocks until unemployment rises or flows pivot, implying a bullish bias with muted upside in the near term. He also discusses energy-dislocation risks, potential price transmission, and why emerging markets are now the marginal oil consumers, shaping how these forces could affect future markets.

Alibaba's Q4 2026 results emphasize an AI + Cloud-driven growth inflection, with revenue up 11% on a like-for-like basis and external Cloud revenue up 40%, while AI-related product revenue accounts for 30% of Cloud external revenue. Management guides model and application services ARR to exceed RMB 10 billion in the June quarter and RMB 30 billion by year-end, propelled by in-house Qwen and AI-native software, along with cloud infrastructure upgrades and MaaS expansion. The company also notes improving consumption metrics, a strong balance sheet with substantial net cash and an annual dividend of USD 1.05 per ADS, and ongoing investments in quick commerce and user experience as part of a long-term value strategy.

The article describes new safety updates for ChatGPT that improve recognition of risk that emerges over the course of sensitive conversations by analyzing subtle cues and combining signals across prior chats, using narrowly scoped safety summaries to inform safer responses. It reports substantial empirical improvements in safe responses in high-risk scenarios (e.g., about 50% improvement in suicide/self-harm cases and 16% in harm-to-others within long single conversations; 52% and 39% improvements on GPT-5.5 Instant across the same categories), as well as high-quality safety summaries (average relevance 4.93/5, factuality 4.34/5) without compromising everyday chats. The work rests on mental health expert collaboration and ongoing research to extend these methods to other high-risk domains.

The post features Boris Mann's assertion that '11 AI agents' is meaningless, because counting agents doesn't reveal their function. He compares it to saying you have '11 spreadsheets' or '11 browser tabs' to do your work, implying that the number of tools is less important than their roles. It offers a concise critique of AI-agent discourse, focusing on tasks and outputs rather than agent counts, rather than providing data-driven analysis.

Seema Amble argues that in an agentic, headless era, defensibility moves from UI-driven stickiness to the data model, permissions, workflow logic, and real-world execution that undergird a system of record. The piece shows how reliance on the UI historically created data hygiene and context, but agents bypass interfaces and expose data directly, making migrations hard due to undocumented SOPs, deep dependencies, and compliance needs. It outlines three strategic paths for buyers—augment incumbents with agent interfaces, DIY rebuild, or buy AI-native replacements—and highlights defensibility criteria for the future, including recreating the SoR, owning proprietary data, closing the action loop, and enabling real-world execution connectivity.