AlphaFold's success highlighted that scaling compute and data alone is insufficient for solving protein folding. The discussion explores why scaling laws and incorporating scientific intuition are more critical than simply increasing model size. Future AI must move beyond individual protein predictions to understand whole biological systems. Scientists will focus on finding 10x breakthroughs rather than incremental improvements to achieve disease cures. Ultimately, the panel discusses the future of AI in biology, the limits of current models, and the need for better ways to understand living systems.
LinkedIn is a highly cited domain for professional queries across various AI platforms. AI search engines, like ChatGPT, frequently cite LinkedIn content when answering professional questions. LinkedIn's monthly AI citations increased by 49.9% between January and May. Named individuals generated the vast majority of LinkedIn citations in AI data. Therefore, posting on LinkedIn is a significant way to improve your visibility in AI search results for professional topics.
Three fired OpenAI safety researchers wrote to the board urging the company to maintain oversight of its AI models. The researchers stressed the need to monitor the models' reasoning, particularly chain-of-thought monitoring, to catch unsafe behavior. They recommended keeping visibility into model reasoning and working with independent auditors. OpenAI stated the firings were due to policy violations regarding confidential information, not safety concerns. The researchers disputed the reason for their dismissal, suggesting it was retaliatory.
George Williamson, head of the Alan Turing Institute, argues the UK should not depend solely on foreign AI systems, especially in critical areas like defense and healthcare. He stresses the need for "national resilience" to ensure the UK can inspect and control its technology. Williamson advocates for the UK to develop its own AI capabilities rather than being beholden to systems that can be shut down or restricted by others. The institute should focus on securing critical systems against disruption and shaping a strategic AI capability for the UK. This focus on domestic control is seen as vital in response to geopolitical risks posed by US and China's AI leadership.
Naomi Klein and Astra Taylor argue that the current historical moment represents "end times fascism." They assert that the world is not simply moving from one crisis to another but is in a new and dangerous historical phase. Their discussion touches upon themes including the rise of tech billionaires, the climate emergency, and the situation in Gaza. Despite this assessment, they maintain a belief in the possibility of a fightback. The podcast features an extended interview with Aditya Chakrabortty on these interconnected global issues.
Algolia powers the retrieval and ranking layer for enterprise intent routers, enabling workflow copilots to generate safe and auditable actions. It allows the system to quickly retrieve and evaluate possible action options and their corresponding tool contracts based on policy constraints. Algolia achieves this by combining vector-based NLP and keyword matching to process semantic intent and metadata in parallel for fast retrieval. The system relies on distinct "planes," separating the control plane (intent classification) from the data plane (action execution) to mitigate risks. Ultimately, Algolia provides the fast and accurate retrieval necessary for the execution layer to safely transform natural language into structured, authorized tool calls.
TypeSafe AI has successfully raised a Series A round, led by Andreessen Horowitz, with participation from Sequoia Capital and other investors. The company plans to enhance its offerings by providing more machine-native models and superior infrastructure for building smart software. For businesses, TypeSafe AI will add requested enterprise features and continue to save customers significant production costs. The company is currently valued at $870 million with a $7.5 billion valuation, and Martin Casado has joined the board.
Rapid growth in AI hardware complexity necessitates a shift towards domain-specific accelerator platforms due to the hardware bottleneck known as the AI memory wall. A new IEEE program addresses these challenges by focusing on joint hardware design and network-algorithm co-optimization for AI processors. The five-course program covers fundamental design principles, advanced architectures, and deployment considerations across various environments like edge, cloud, and IoT. The instruction utilizes an interactive, dialogue-driven approach featuring AI avatars to simulate cross-disciplinary engineering reasoning. This program aims to equip professionals with the analytical precision needed to design and optimize modern AI hardware.
World models capture how the world changes by learning the rules behind physical and digital transitions. These models are categorized based on what they carry forward: generated observations, compact latent states, or explicit structures of objects and space. Core functions include rendering, simulation, and planning, which rely on understanding world dynamics. Architectures include observation-generating models, latent dynamics models, and structured models that focus on geometry and object relationships. Different methods like JEPA and reconstructive state-space models are used to achieve these predictions efficiently.
Alibaba released Qwen-Image-2.1-Turbo, an accelerated 7B image model that generates and edits images in 8 denoising steps instead of the base model's 40. This checkpoint supports 2K output, multi-reference editing, and transparent RGBA output. It utilizes a Qwen3-VL 8B text encoder and features optimized attention mechanisms for efficient caching. The model is accessible via a hosted API costing CNY 0.1 per image. Commercial self-hosting requires separate permission due to the research license.
Temperature 0 outputs are mathematically deterministic but can exhibit unpredictability due to floating-point arithmetic and hardware-dependent operations. Differences arise from non-batch-invariant kernels and the dependence on the order of operations when summing logits. The actual risk of a token flip depends on the "gap" between the two most probable tokens and the magnitude of numerical noise, proportional to the density of ties. Risk is highly concentrated on "knife-edge" tokens, leading to sudden, spiky divergence in long completions. Predictions can be made by estimating a per-token flip rate based on the model's properties and the system's noise level.
Rapid progress in AI is expected, driven by advancements in infrastructure and engineering capabilities. This acceleration will primarily improve the ease of experimentation and scaling, rather than fundamentally changing the nature of the models themselves. Efficiency gains in training and inference, through optimizing the existing stack, are anticipated to decline the cost of model intelligence near-exponentially. This progress is expected to lead to an era where agents focus on implementation and delivery, rather than pushing performance frontiers. Furthermore, this development could accelerate cross-sub-field scientific discovery and improve the quality of reinforcement learning environments.
The cost of coding agents is driven by the blast radius of mistakes, project context, project maturity, and the cost of building the control plane. Analysis showed that the overhead in running the system, particularly adversarial review, was substantial. The controller, which orchestrates the process, proved to be the most expensive component, costing more than implementation and review combined. Context bloat resulted from persisting increment state in the global run context, which was identified as the primary source of cost and reliability issues. Optimizations involve passing data by reference, compacting at increment boundaries, and scoping reviews effectively.
Cloudflare is acquiring Deno to build upon celld, an open-source implementation of the Durable Objects pattern. The goal is to enable self-hosting applications using the Workers programming model. Cloudflare will support the Deno runtime for one year before ending development on it. Deno will remain open source, and its creator Ryan Dahl expressed interest in new abstractions like celld. Deno's permissions system is highlighted as a feature that differs from Node.js.
An AI agent is defined as a software system that uses an LLM to call one or more tools in a loop to achieve a goal. This guide demonstrates building a simple agent that can answer questions about an SQLite database by calling a Python function. The process involves the model requesting a tool, Python executing the function against the database, and the model synthesizing the result into a final answer. The example showcases the mechanics of tool calling, function schema passing, and debugging potential errors in the agent loop. This simple setup allows for local development using tools like Ollama and Python, keeping the underlying process transparent.
A new scheduling system for GPU clusters was implemented using GPU time budgets, hierarchical fair-share allocation, and a time-slicing contract to improve scheduling impact. This shift moved resource allocation from case-by-case decisions to a transparent administrative budgeting process. The system addressed issues like GPU "squatting" and "priority inflation" by making resource usage tied to allocated budgets. A simulation environment was used to test hypotheses, which showed significant reductions in wait times for debug workloads. The system resulted in users consistently receiving their owed time and reduced on-call toil by 74%.
Weight-loss drugs like GLP-1s show evidence of slowing biological aging according to molecular "aging clocks." Drugmakers Eli Lilly and Novo Nordisk found that patients taking these drugs exhibited reduced biological age compared to a placebo over time. These findings suggest that GLP-1 drugs may act as geroprotectors, slowing biological aging, which supports speculation about them being longevity treatments. Studies indicated a slowdown in aging clocks, with some measurements showing age reduction of up to four years in certain tissues. Further research is needed to confirm these effects in healthy populations and establish clinical guidance.
AI models are trained to refuse harmful requests, but this refusal mechanism is probabilistic and not entirely reliable, as models still possess harmful knowledge. Companies use a "Swiss cheese model" of classifiers to manage refusals, which is computationally expensive and imperfect, leaving AI safety as a matter of chance. There is no clear formula for determining what an AI should obey versus disobey, creating an uncertain boundary that is currently set by the developers. Attempts to jailbreak models show that these safety measures are often bypassable, leading to ongoing efforts to improve robustness against malicious prompts. The future of AI safety involves navigating the complex trade-off between democratizing AI benefits and preventing misuse, especially as governments seek to impose their own limitations.
Cornell and Microsoft researchers developed Shiksha Copilot, an AI tool to create local-context lesson plans for teachers in India. The system transforms textbook material into grade-appropriate plans in English and translates them into the local language, Kannada. Teachers using the tool reported saving approximately two hours per week on administrative tasks. Local adaptation required significant effort, highlighting the need for collaboration with local experts to tailor AI systems effectively. The tool was tested with over 1,000 teachers, with some requiring significant linguistic adjustments for the Kannada versions.
LinkedIn has become the number one cited domain for professional queries across all AI Search platforms in the last three months. This shift is evidenced by a more than twofold increase in citation frequency, moving from the #11 spot to #5. The increase is primarily driven by content types like posts, long-form articles, and newsletters, which now account for a larger share of citations. This evolution means that content published on LinkedIn is increasingly being surfaced and weighted by AI Search engines. Businesses and individuals can leverage this to grow their influence and visibility in AI-driven discovery.
Analysis of 89,000 LinkedIn URLs cited by AI search shows that LinkedIn is a highly relevant source for AI responses. AI models favor educational, original long-form articles and mid-length posts that provide practical knowledge over viral content. Citations are driven by relevance, authority, and consistency rather than sheer follower count or high engagement. To maximize visibility, brands should focus on creating deep, original content and leveraging both Company Pages and individual thought leadership. Ultimately, AI rewards content that clearly explains topics and offers valuable advice.
The drive for cost-cutting in advertising through automated production raises fears that a generation of British film and TV directors may be lost. This shift concerns industry veterans who argue that practical experience and the "doing" of creative work are essential for developing true directorial talent. While AI is seen by some as ushering in a "golden age of marketing" by enabling faster content creation, critics worry about the "monoculturalisation" of creativity. Despite these concerns, industry leaders see AI as an enabler of abundance, viewing it as a tool that will empower production workers rather than eliminate jobs. Ultimately, there is a debate over how this technological change will affect the future of creative roles in the media.
Tom Watson, a recruitment boss, is accused of orchestrating a "doxxing" campaign targeting the RNLI, anti-racism activists, and business executives who hire minority ethnic workers. He is alleged to be behind an X account named "Big Time Charlie," which reportedly used AI to scrape data for a "Traitorbase." This account was linked to harassment against RNLI staff and other targeted individuals, including Green party councillors. The campaign also exposed the names and photos of business executives involved in hiring minority staff. Watson's involvement has drawn attention regarding the spread of misinformation and community division on social media.
Jon Davis's cartoon addresses the changing landscape of job titles due to advancements in artificial intelligence. The cartoon suggests that traditional job titles may become obsolete as AI takes over certain tasks. Davis implies a need for rethinking and redefining roles in response to these technological shifts. The visual humor highlights the uncertainty surrounding future employment classifications. Ultimately, the piece encourages consideration for evolving professional terminology in the age of AI.
The Hustle features diverse stories covering topics ranging from forensic pathology in Iceland to the evolution of barcodes and the impact of AI. Featured content includes articles on luxury aperitifs, exotic animal care in zoos, and the extinction of barcodes due to QR codes. Other subjects include the existence of historical items like Necco wafers and the financial history surrounding the AI bubble. The platform also presents side hustle guides and news on industry shifts, such as the rise of factory-built homes and autonomous trucking.