Full-text search across article titles and summaries. Refine with sector, year, content type, and reading-time filters.
Anthropic's third economic index reveals that Claude usage is heavily concentrated in wealthy countries like Israel and Singapore, with businesses using AI for automation 77% of the time compared to consumers' 49%.
The system prompt for OpenAI's Codex has been leaked.
Sponsored
As the next generation of AI applications comes to market, ambitious teams are turning to Baseten for inference infra [4] that not only keeps up with but accelerates their innovation.
xAI has reportedly laid off a third of its data annotation team as it pivots to expand its specialized AI tutor division.
Grok 4 Fast, the newest addition to xAI's lineup, is now available for users on the Grok web interface via the model selector. It can be accessed by enabling a new toggle in the Subscription settings. Marked as an early access beta, Grok 4 Fast is up to 10 times quicker than the standard Grok 4. It is optimized to respond rapidly by spending minimal processing time on complex tasks, which limits its creative abilities.
A walkthrough of the entire post-training lifecycle of LLMs, from supervised fine-tuning and reward modeling to reinforcement learning methods such as RLHF, along with evaluation best practices.
Despite billions invested, 42% of enterprise AI initiatives were discontinued in 2024. This was caused by how models were embedded into business. The winners redesign workflows, rethink structures, and take ownership of the service layer where value is created. The next generation of CEOs will treat AI as a labor class and deploy the technology with the same discipline that the most successful serial acquirers apply to capital.
This article evaluates different jailbreaking methods against gpt-oss. The model appears to have had robust safety training in both system prompting and refusal vector attacks. It is tricky to work with, and not all libraries support its idiosyncrasies.
Sponsored
The 2025 Gartner® Magic Quadrant™ for Strategic Cloud Platform Services [10] compares Google, Microsoft, Amazon, and 5 other leaders in cloud infrastructure. It looks at dozens of mandatory and common features that determine a vendor's ability to support mission-critical
VaultGemma is a model that Google trained from scratch with Differential Privacy (DP). DP offers a mathematically robust solution to user privacy that adds calibrated noise to present memorization. It has some trade-offs, like reducing training stability and significantly increasing batch size. There is still a utility gap between DP-trained and non-DP-trained models, but that gap can be systematically narrowed with more research on mechanism design for DP training.
Real-world value often stems from the length of a task an agent can complete. Marginal gains in single-step accuracy can compound into exponential improvements in the length of a task a model can successfully complete. Models are more likely to make mistakes when the context contains errors from previous turns. Failures when tasks are made longer arise from mistakes in execution rather than an inability to reason.
Teams that shift from API shaped tools to workflow-shaped tools see meaningful improvements in reliability and efficiency. MCP works best when tools handle complete user intentions rather than exposing individual API operations. Large language models don't work like developers - they have to constantly rediscover which tools exist, how to use them, and in what order, so building tools around workflows produces better results.
Models aren't perfectly interchangeable - if you are switching models, rewrite your prompts. Prompts overfit to models the same way models overfit to data. They need to be tested, evaluated, and aligned with the defaults of the new model. Adapting prompts will save tokens while producing better results.
Most of the new value created by AI will be captured by consumers, who will see wider and more affordable access to services like medical care, education, and advice. Knowledge-intensive services will get cheaper, allowing consumers to buy more of them. At the same time, services that require person-to-person interaction will get more expensive and take up a greater percentage of household spending. There will be obvious opportunities in both. Think through the implications of knowledge workers becoming more efficient, imagine what markets this efficiency unlocks, and invest in those.
Sponsored
Warp already beat Claude Code and Cursor in agent benchmarks. Now it has a nifty editor, code review, and other tools that make it the
Cornell's Center for Advanced Computing published an interactive workshop covering GPU memory hierarchies, streaming multiprocessors, and detailed breakdowns of NVIDIA's Tesla V100 and Quadro RTX 5000 architectures.
How to manage short-term memory for AI agents using the OpenAI Agents SDK, employing trimming and compression techniques to keep sessions coherent, fast, and reliable.
Nvidia now uses its DGX Cloud capacity for internal research.
OpenAI has announced a 5-week residency for early-stage technical founders, offering mentorship, early tool access, and peer collaboration to explore new AI product ideas.
Sponsored
It's unrealistic to completely avoid AI content in 2025. But without quality guardrails in place, it can end up surfacing irrelevant or poor quality results in site search — and badly frustrating users.
Microsoft and OpenAI will extend their partnership. The move could help OpenAI move towards becoming a for-profit corporation. OpenAI plans to create a new for-profit company in which Microsoft and OpenAI will each initially receive a roughly 30% stake. The rest will go to employees and investors. It plans to keep the nonprofit's control over the new for-profit and endow it with a stake valued at more than $100 billion.
Qwen3-Next is a new model architecture featuring a sparse Mixture-of-Experts design with hybrid attention and multi-token prediction. Its 80B-parameter base model activates only 3B parameters at inference, enabling 10x faster throughput for long-context tasks.
Claude starts conversations with a blank slate and only activates memory when users explicitly invoke it. It recalls by only referring to users' raw conversation history. ChatGPT's memory system exposes interaction metadata, recent conversation context, model set context, and user knowledge memories alongside the system prompt. That the two top assistants have built completely opposing memory systems shows that memory in AI has a massive design space with no right answer or one-size-fits-all technique.
Anthropic engineers found that agents perform better with fewer, more thoughtful tools rather than wrapping every API endpoint. Claude-optimized tools significantly outperformed human-written versions in internal tests, with agents able to automatically improve their own toolsets through evaluation loops.
Sponsored
The AI Security Summit [9] will be held in person from October 22, 2025 - October 23, 2025! Join an exclusive gathering of AI innovators and execs confronting the AI Security Chasm—and building a foundation of trust for future-forward AI initiatives.
Infrastructure choices matter a lot for distributed training. Network and storage configurations can easily create 6-7x performance differences. This post looks at network storage and configurations that directly impact both training time and costs.
OpenAI's recently released GPT-OSS series of models features novel techniques like MXFP4 quantization, efficient kernels, a brand-new chat format, and more. The library has been upgraded considerably. This post looks at all of the upgrades in-depth. The updates make it very efficient to load, run, and fine-tune the models.
Cursor Tab is a system that predicts users' next actions across their codebase to make suggestions. The Tab model runs on every user action, handing over 400 million requests per day. This has given Cursor a lot of data about which suggestions users accept and reject. Cursor uses this data to improve Tab using online reinforcement learning. Its approach involves rolling out new models to users frequently throughout the day and using that data for training.
GPT-5's launch was so botched that Sam Altman apologized for it. However, it appears business adoption has continued to increase after a slowdown earlier in the summer. It looks like GPT-5's release drove a sharp rebound for OpenAI's tools. Adoption growth is outpacing Anthropic for the first time since May.
Meta's TBD Lab operates in a special badge-access area near Zuckerberg's desk, hidden from internal org charts. New hires routinely leverage competing offers from other frontier labs to secure massive compensation packages. ChatGPT co-creator Shengjia Zhao quit within a week of joining but returned after Meta tripled his pay and made him chief scientist. Meanwhile, existing employees lobby for raises while being excluded from the elite unit's secretive operations.
Sponsored
Rovo is Atlassian's AI-powered core, helping every team work smarter and faster with AI-powered search, chat agents, and more. Unlike alternatives, it comes with the security, compliance, and guardrails
Switzerland's open-source AI model, Apertus, is an alternative to proprietary models like ChatGPT.
This article outlines three key perspectives on AI's future, focusing on its development trajectory and societal impact.
Perplexity has secured another $200 million in funding, pushing its valuation to $20 billion.
Sponsored
Security?
OpenAI reportedly signed a $300 billion deal with Oracle for compute power over five years, starting in 2027, marking one of the largest cloud contracts ever. This move diversifies OpenAI's cloud providers beyond Microsoft Azure, aligning with its involvement in the $500 billion Stargate Project for domestic data centers. Oracle has been working with OpenAI since 2024, highlighting a strategic shift in cloud resources essential for OpenAI's growth.
The AI coding platform grew annualized revenue from $2.8 million to $150 million in less than a year, driven by 40 million users who can now build apps without coding experience. Agent 3 can work autonomously for up to 200 minutes compared to Agent 1's 2-minute limit, testing and fixing its own code while building applications with minimal human interaction.
ByteDance's Seedream 4.0 is an AI image model that the company claims is superior to Google DeepMind's "Nano Banana" in prompt adherence, alignment, and aesthetics. The model combines features from previous versions at the same cost of US$30 per 1,000 generations. These claims are based on ByteDance's internal MagicBench benchmark but lack an official technical report.
LLM inference remains nondeterministic even under greedy decoding due to floating-point non-associativity and concurrent execution.
It's difficult to understand how model providers will manage to generate sufficient usage, as the adoption rate is far from sufficient to absorb the computing power that will eventually be brought to market. The semiconductor industry will have to enter a downward cycle at some point, and this will hurt the AI sector significantly. The lack of interest in the economics of semiconductors is a real weakness in the analysis of the development of generative AI. This article shares economic considerations on the subject to help readers better understand these contemporary issues.
What if AI follows predictable patterns of adoption instead of the rapid superintelligence takeoff many in Silicon Valley prophesy? The impacts of AI depend on how it's adopted, not raw capabilities alone. Many industries continue to use outdated technology due to the non-technical (for example, Air Traffic control uses the same system from the 50s). Even advanced AI features available today have an adoption problem. ChatGPT's "thinking" models are used by less than 1% of users nearly a year after release.
Sponsored
AI at work is growing fast — Slack's latest Workforce Lab report shows daily use up 233% in six months [11], driving 64% higher productivity and 81% greater job satisfaction. Join Slack's webinar [11] to dig into the data and learn practical ways to use AI for smarter insights, faster workflows, and less busywork.
The new web fetch tool in the Claude API fetches content from URLs as part of its response to prompts. It can extract full text content from URLs and PDFs. For security, the tool can only fetch URLs that have previously appeared in the conversation context, so it cannot fetch arbitrary URLs that Claude generates or URLs from container-based server tools. Anthropic recommends only using the tool in trusted environments or when handling non-sensitive data, as enabling it poses data exfiltration risks.