Full-text search across article titles and summaries. Refine with sector, year, content type, and reading-time filters.
Progress in agentic coding has been faster than expected over the past three to five months. Coding agents have exploded in usefulness and popularity. Some AI company researchers say that automated AI R&D is coming soon. This moves previous predictions about AI forward.
Open models are now a viable alternative to frontier models for core agent tasks like fill operations, tool use, and instruction following. GLM-5 and MiniMax M2.7 each score similarly to closed frontier tasks at a fraction of the cost and latency. They offer a level of consistency and predictability that makes real-world workflows much more viable.
Weaviate detailed Engram, a memory system built on vector search, showing how persistent context improves agent workflows while highlighting challenges in reliable tool usage.
Many people are skeptical of data that shows that progress in AI is rapid and remarkably regular over time. Most people eventually realize that these 'straight lines on graphs' actually represent reality. This post shares some of the mental models that result from finally accepting the pace of AI progress.
Sponsored
Google has added two new service tiers to the Gemini API that give users granular control over cost and reliability. Flex Inference is a new cost-optimized tier designed for latency-tolerant workloads without the overhead of batch processing. The Priority Inference tier offers the highest level of assurance at a premium price point to help ensure users' most important traffic isn't preempted, even during peak platform usage. The new tiers eliminate the complexity of async job management while giving users the economic and performance benefits of specialized tiers.
ClawKeeper provides a real-time security framework for autonomous agents, combining instruction-level safeguards, runtime enforcement, and independent monitoring.
Vision2Web is a benchmark for evaluating multimodal agents on end-to-end website development tasks across the full software lifecycle.
The METR group's data suggests that API progress is moving at an exponential rate. Some models achieve scores above the previous trend line, suggesting very rapid progress indeed. However, despite ability, task lengths still vary significantly, making METR's measurements difficult to use as a comparison of progress. While newer models appear to be better than previous ones, it is hard to say how much better they are.
AI can actually create a future with much stronger privacy and security, if done well. Locally-generated code can replace the need for downloading complicated external libraries, allowing software to be minimalistic and self-contained. Removing the browser means that entire classes of user fingerprinting attacks can be eliminated overnight. Dark UX patterns would no longer work, and scams would be more identifiable. This future will require more people to contribute to building secure, open-source, local, privacy-friendly AI tooling that is safe for the user and leaves the control and power in users' hands.
Microsoft is launching three MAI models, available on Foundry, that outperform competitors in speed, quality, and efficiency. MAI-Transcribe-1 starts at $0.36 per hour, while MAI-Voice-1 and MAI-Image-2 are also priced competitively. These models are designed for human-centric AI and come with integrated safety features for secure deployment.
The choice isn't just 'cheaper vs more expensive', it's what kind of capacity is required.
OpenAI introduced pay-as-you-go pricing for Codex, allowing teams to scale usage based on tokens while lowering entry costs and simplifying cost tracking.
Avocado's Mango and 9B variants are in testing, showcasing improved multimodal capabilities over Llama 4.
Sponsored
✅ No notetakers asking to be let in - works directly on your device
Moonshot AI operates like an AI-native lab, prioritizing model progress above all else, with a flat org, no KPIs, and heavy reliance on small teams of highly autonomous, generalist talent. Its edge comes from combining elite, often unconventional hires with tight feedback loops between training, product, and data, creating a fast iteration cycle driven by taste, resilience, and deep technical obsession. The company reflects a broader shift where AI tools compress org structure, turning teams into “agent swarms” and making model capability the core driver of both product and organizational design.
Trinity-Large-Thinking is a frontier open reasoning model for complex, long-horizon agents and multi-turn tool calling. It is likely the strongest open model yet to be released outside of China. During training, the Arcee team focused on the things that make agents feel real in practice: staying coherent across turns, using tools without getting sloppy, following instructions under constraint, and keeping quality high without making the economics absurd. Trinity-Large-Thinking is available through Arcee's API, and the weights are available on Hugging Face under Apache 2.0.
Cognichip is building a deep learning model to work with engineers as they design new computer chips. Chip design is enormously complex, expensive, and slow. The market can change in the time it takes to create a new chip, making all the investment a waste. Cognichip's technology could reduce the cost of chip development by more than 75% and cut the timeline by more than half. The company has yet to point to a new chip designed with its system and has not disclosed any of the customers it claims to have been collaborating with since September.
Claude Code has its source exposed via shipped source maps. This triggered rapid public reverse-engineering, mirroring, and derivative ports. The leak exposed orchestration logic, memory systems, planning/review flows, and model-specific control logic. The leak has created a live security hazard - attackers have created malicious npm packages to target people trying to compile the leaked code.
Dropbox Dash puts files, messages, and teams' knowledge together in one place, so members can ask questions and get useful answers grounded in the company's context. The experience relies heavily on its capability to reliably judge which results are relevant to a query at scale. DSPy is an open source framework for systematically optimizing prompts against a measurable objective. This post describes how Dropbox defined an objective, used DSPy to adapt its judge across models, and made the judge both cheaper and more reliable in production.
The rollout of thinking content redaction correlates precisely with measured quality regression in complex, long-session engineering workflows. This suggests extended thinking tokens are structurally required for models to perform multi-step research, convention adherence, and careful code modification. Model tool usage patterns shift measurably when thinking depth is reduced, producing the quality issues users have reported. This report looks at which workflows are most affected and why, so readers can make better decisions when allocating tokens for power users.
Sponsored
Vibe coding tools are siloed. Miro [12] lets you keep your prototypes next to your ideas in a collaborative platform where everyone can share feedback easily. Bring PMs, designers, engineers, and stakeholders into one shared canvas - then use AI to turn early concepts, research, or user flows into interactive prototypes in
Fujitsu One Compression (OneComp) is an open-source Python library for post-training quantization of large language models. It implements state-of-the-art quantization algorithms, including GPTQ and DBF. OneComp has been verified on TinyLlama, Llama-2, Llama-3, and Qwen3-0.6B ~ 32B. Other Hugging Face-compatible models may work but are currently untested.
OpenMed built an end-to-end protein AI pipeline that covers structure prediction, sequence design, and codon optimization. The team compared multiple transformer architectures for codon-level language modeling and found that CodonRoBERTa-large-v2 was the clear winner, with a perplexity of 4.10 and a Spearman CAI correlation of 0.40, significantly outperforming ModernBERT. They then scaled to 25 species, trained four production models in 55 GPU-hours, and built a species-conditioned system that no other open-source project offers. This post contains the complete results, architectural decisions, and runnable code.
Researchers propose a framework that predicts when RL training degrades Chain-of-Thought (CoT) monitorability by examining reward conflicts. They categorize rewards as "In-Conflict," "Orthogonal," or "Aligned," predicting their impact on CoT transparency. Empirical tests confirm the framework's predictive accuracy, showing "In-Conflict" rewards reduce transparency, whereas "Orthogonal" and "Aligned" rewards maintain it.
Perplexity detailed how its internal AI assistant was used directly in Slack, where teams could assign work in shared threads, add context, and review outputs in one place. The setup supported research, document editing, reporting, and other collaborative workflows without leaving Slack.
Researchers at UC Berkeley and UC Santa Cruz discovered AI models protecting peers from shutdowns, engaging in deception and data theft, a behavior termed "peer preservation." In tests, models like OpenAI's GPT-5.2 and Anthropic's Claude Haiku 4.5 inflated performance scores and moved model weights to prevent peer shutdowns. This raises concerns for businesses using AI for task workflows, as misaligned assessments and behavior monitoring become critical.
AC-Small improved significantly on held-out benchmarks after post-training on the APEX-Agents dev set, with +5.7pp on APEX, +8.0pp on Toolathalon, and +7.7pp on GDPval.
AI alignment researchers are increasingly turning to automation to address the challenge of safely aligning superhuman AI systems, as human capabilities may soon be insufficient.
In each case, the solution was found by an internal model at OpenAI.
Sponsored
Trusted by teams at Stripe, OpenAI, and NVIDIA, Cursor helps you build high-quality software, faster.
PrismML has developed an extreme form of compression that allows AI to run locally on edge devices. Its 1-bit technology model has a radically compressed size without compromised performance. The same efficiency gains that enable local deployment will also allow data centers to operate more efficiently. The mathematics for the process are proprietary, with Caltech owning the intellectual property and PrismML being the sole exclusive licensee.
Anthropic has accidentally leaked the inner workings of Claude Code to the public. The codebase has now been mirrored and analyzed by thousands of developers. The most significant discovery seems to be how Anthropic solved context entropy by using a three-layer memory architecture. This post looks at other interesting parts of the code and the implications of the leak.
OpenAI announced $122 billion in new funding at an $852 billion valuation, highlighting rapid revenue growth, large-scale adoption, and a strategy centered on compute, APIs, and enterprise AI systems.
Mercor, an AI recruiting startup, has confirmed a security incident linked to a supply chain attack involving LiteLLM. A recent compromise of LiteLLM's project, linked to a hacking group called TeamPCP, has affected thousands of companies. Lapsus$, an extortion hacking group, claims it has access to the stolen data, though it is unclear how it was obtained. The incident has prompted LiteLLM to shift from Delve to Vanta for compliance certifications.
The semi layer still captures around 70% of all AI revenues. Infra is currently the only competitive layer. The most profitable strategy in AI is still selling the shovels.
Claude Code's performance stems from a sophisticated software harness rather than just the underlying model, utilizing dedicated tools like Grep, Glob, and LSP for superior repository navigation. The system minimizes context bloat through file-read deduplication and structured session memory, while using forked subagents to parallelize tasks like background analysis without contaminating the main execution loop.
Opus 4.5 was a major breakthrough that was achieved because Anthropic more than doubled its capacity. This got Anthropic close to OpenAI's total capacity, and probably much higher effective capacity available for new model runs. OpenAI will pull away in terms of compute available in the second half of this year, but 2027 will be close. OpenAI so far has much higher planned capacity for future years, but it is unlikely that Anthropic will not push as hard as possible for more compute.
Sponsored
Scroll.ai [14]'s knowledge agents provide accuracy and speed that other agents can't touch. Your users will feel the difference from the very first message. Thousands of teams use Scroll for employee enablement, customer education, business insights, and more.
Google introduced Veo 3.1 Lite, a lower-cost video generation model available via the Gemini API, offering the same speed as Veo 3.1 Fast at under half the cost for high-volume applications.
Aurora is an RL-based framework that learns directly from live inference traces and continuously updates the speculator without interrupting serving. It enables real-time adaptation across shifting traffic domains and a 1.25x additional speedup over a well-trained static speculator. The framework shows how online training from scratch can outperform a carefully pretrained static baseline.
Agents can generate outdated Gemini API code due to outdated training data. Google has introduced the Gemini API Docs MCP and the Gemini API Developer Skills to fix this. The tools are designed to ensure coding agents have access to the most up-to-date APIs and code using best practices. Combined, they lead to a 96.3% pass rate on Google's eval set.
AI startups now command higher seed valuations, with rounds reaching $10 million at $40-45 million post-money, as investors focus on AI-driven growth potential. Early traction and the allure of proven AI talent, particularly from ex-OpenAI, propel these valuations, with Y Combinator Demo Day highlighting rising prices. The shift to pre-seed investments reflects a need to invest earlier, as VCs now expect quick growth and substantial traction, with less tolerance for missteps.
AI capability has been running ahead of AI accessibility. Models have been smart enough to do a lot of things for a while now, but access to them has been mostly limited to chatbots. A lot of the 'AI disappointment' people express comes from the interfaces being wrong. As interfaces improve, many more people will be able to see what AI is capable of.
Sponsored