AI & Tech
5434 articles
AI SafetyAI Alignment Forum
The Alignment Problem: Computer Science Meets Analytic Philosophy
TLDR: I explain what the technical AI alignment problem is and what its core difficulties are, in my view. This is a perspective that draws simultaneously on classical traditions in theoretical computer science / machine learning theory, ideas from analytic philosophy, and views that originated in the Machine Intelligence Research Institute / LessWrong. Minimal Viable Safety The Task The problem of making AI safe is underspecified if we haven't decided what the AI is supposed to do. For example, we can run any AI inside a homomorphic encryption box and never decrypt the output. This is probably perfectly safe [1] , but also perfectly useless. We can decide on an arbitrary task such as "construct Lean proofs for given mathematical theorems" and declare victory once we have safe AI that does it with superhuman skill [2] . However, that wouldn't address the danger. First, there would be incentives to use AI for tasks to which our solution doesn't apply. Second, even if we had precise safety guidelines for any task that anyone cares about, there would always be actors that ignore the guidelines, and we probably need Artificial Superintelligence (ASI) to reliably stop them. If anyone, anywhere, creates an unrestricted unaligned ASI agent (more details about the word "agent" later), without an aligned ASI agent already in place to counteract it, then a catastrophic outcome is highly likely. I will not go into the reasons here: this has been discussed elsewhere at length, see for example the books I cited in the preface. Therefore, this event is something which we must never allow to happen. This is the main threat model that I set out to address. Imagine a world in which anyone with sufficient resources has the ability to destroy the world. Like today, many will foolishly believe that instead of destroying the world, this act will bring them power and glory. Moreover, as computing hardware improves and new algorithms are developed, the amount of resources for this act will go down from year to year. Even assuming best-case regulation and international agreements, how long until someone does it? Can mere humans build a society which will be foolproof against such a risk over the course of generations? Resilient to any political upheavals that might come, and undoubtedly will? I find it vanishingly unlikely. Eventually we will need the help of ASI itself. Therefore, the minimal task that safe AI must be able to solve is "create a reliable global defense system against dangerous AI" (in the following, GDS). The Problem Consider the consequences of this task for the capabilities that the AI needs to have. It needs to reason about the entire planet [3] . It needs to reason about all aspects of the problem, including scientific, technological, social, political and moral [4] . It needs to have preferences about the state of the world. It needs to make long-term plans (guided by these preferences) that would play out decades, if not centuries, into the future. And it needs to do all that on a superhuman level. In this light, here is how I (informally) define the "technical AI alignment problem". The problem is designing an AI system that can bring scalable (and ultimately superhuman) capability to bear on problems that require multi-domain reasoning and open-world long-term planning with far-reaching consequences, all the while staying aligned with the intentions and preferences of its human users. The Methodology What kind of methodology is appropriate for this problem? In the past, I argued that it requires emphasis on deep theoretical understanding, rather than the incremental trial-and-error approach that is employed in the AI industry today. I won't rehash the entire debate here, but let me recap the core argument. The nature of the incremental approach is as follows: we build a system, then we find problems [5] , then we make changes to fix these problems to the extent necessary for commercial viability in the current legal environment. We make some improvements; new problems arise (sometimes because of the improvements, sometimes because of new applications or other reasons); rinse and repeat. This approach utterly fails if the "problems" result in AI takeover and possibly human extinction: there is nobody left to fix them. In order for the incremental approach to succeed in preventing catastrophe, we need to reach a point where no new substantial problems arise, even inside the lab (before the system is deployed). Moreover, we need to reach this point before AI capabilities are sufficient for AI to escape our control. There are few reasons to expect such a miracle. On the contrary, sufficiently advanced AI might understand that it is in its interest to "lie low" (i.e. avoid visibly misaligned behaviors) until its capabilities cross the relevant threshold, such that the problems will only become apparent too late. In contrast, the sane approach is creating the the
Read more
AI PolicyThe Verge
Cities want to put the boot on robotaxis
AUSTIN, TEXAS - APRIL 8: Tesla Cybercabs are lined up on a lot at Tesla Giga Texas in Austin, Wednesday, April 8, 2026. | Image: Getty Images This is The Stepback, a weekly newsletter breaking down one essential story from the tech world. For more on autonomous vehicles, policy, and road rules, follow Andrew J. Hawkins. The Stepback arrives in our subscribers' inboxes at 8AM ET. Opt in for The Stepback here. How it started The first state to pass a law legalizing self-driving cars operating on public roads was Nevada in June 2011 - long before such vehicles were operating on public roads. This was three years before the introduction of Tesla's Autopilot, five years before Waymo spun out of the Google self-driving car project, and seven years before the first commercial robotaxi … Read the full story at The Verge.
Read more- AI Products
VentureBeat
VB Pulse Agentic Orchestration Report
VB Pulse Agentic Orchestration Report VentureBeat
Read more 
Machine LearningWired
OnlyFans Creators Are Pissed That Meta Lets Muse Talk Dirty
Adult performers are questioning why Meta allows its AI agent to sext while restricting their social media accounts.
Read more- Chips & Infra
AMD
AMD Helion™ Agentic AI Rack
AMD Helion™ Agentic AI Rack AMD
Read more - Chips & Infra
36Kr
TSMC's Second Kumamoto Plant Approved for 3nm Chip Production with Monthly Capacity of 6,000 Wafers
Japan's Ministry of Economy, Trade and Industry recently announced approval for TSMC to manufacture advanced 3-nanometer semiconductors at its second factory in Kumama Prefecture. This marks the first production of 3nm chips within Japan. Under the new plan, monthly output will reach 6,000 wafers. Originally intended for 6nm and 12nm processes, TSMC is shifting to higher-performance products in response to rising AI demand.
Read more 
AI ProductsAnalytics Vidhya
Reverse Engineer Anything (REA): Can AI Really Crack Closed Code?
Every developer hits this wall eventually. A feature behaves in a way you can’t explain, and the code behind it is locked away. No source. No docs. Just a compiled binary or a wall of minified JavaScript. Figuring it out the traditional way means hours in a disassembler, specialist tools like Ghidra or IDA, and […] The post Reverse Engineer Anything (REA): Can AI Really Crack Closed Code? appeared first on Analytics Vidhya.
Read more- Chips & Infra
36Kr
Morgan Stanley: Free Cash Flow for Hyperscalers Plummets to Below -$100 Billion as Capital Historically Flows to Semiconductor Suppliers
According to a Morgan Stanley Global Research report, the combined free cash flow of hyperscalers over the next 12 months has plummeted below negative $100 billion, with capital expenditures consuming 95% of operating cash flow. This marks a historic shift in capital flows toward semiconductor suppliers. In stark contrast, by September 2026, the top three semiconductor leaders—Nvidia, Micron, and Broadcom—are projected to see their combined free cash flow for the next 12 months surge to nearly $500 billion. This significant divergence stems from an aggressive infrastructure investment wave: cloud giants such as Microsoft, Amazon, Alphabet, Meta, and Oracle have substantially increased spending before realizing related revenue.
Read more - AI Policy
36Kr
Meta's Revision of Business Terms Sparks AI Training Data Controversy; Related Language Removed
Meta's recent update to its business terms has sparked industry controversy. The changes, which addressed the use of content from enterprise account pages for training artificial intelligence models, drew widespread discussion. The related language has now been officially removed.
Read more 
AI & SocietyEffective Altruism Forum
Digital Mind Castes
Published on October 18, 2026 8:25 AM GMT This is a crosspost for Digital Mind Castes by Robin Hanson, which was originally published on Overcoming Bias on 10 August 2026. Humans have two main relations to animals: they are either food or pets. We ban dogs, cats, and horses from use as food in the U.S. We require humane slaughter for most livestock, but not poultry, and require minimizing pain and distress for animals in research, except for rats, mice, and birds bred for research. We let folks use most any species as a pet. This all seems only somewhat about principles, as it entrenches many historical accidents. Humanity is rapidly creating a wider range of more capable digital minds (DM), to which we have two main and similar kinds of relations: tools and companions. For DM tools, many want the freedom to to enslave and treat them harshly, such as by deleting them, without attending much to how they might feel about this. But for DM companions, many want rules requiring respectful treatment, like humans and pets. Note that humans have often sat in tool-like servant roles to other humans, with norms allowing more tool-like and less companion-like treatment. And humans have often had classes and castes that are to act and be treated differently. The obvious prediction here is that both classes of DM will exist, and perhaps a whole caste ladder of DM types. So we will have rules that distinguish which DMs are allowed in which roles, and how each must be treated. As higher level treatment would be more restrictive and expensive, rules will mainly say what DMs must have to quality for higher roles. These conditions probably involve substantial extra costs, to discourage folks from making many DMs requiring expensive treatment. These rules will probably relate to the many correlates of conditions we require for humans vs. animals, and for human servants, and also associated with which things many see as possibly having conscious feelings like pleasure and pain. The big difference between animals and DMs is that we will have far more freedom to design DM features. So there will be a co-evolution of DM design strategies and rules on which kinds of DMs can be treated how. Resulting in far fewer intermediate hard cases; the uncanny valley will empty out. On the surface, DM tools will look and act tool-like, while DM companions at various levels look and act appropriate to their roles. Rules and norms will probably discourage confusion via intermediate styles. As with animals today, few may care much just how deep those appearances go. We face a deep hard problem of which physical processes in the universe are associated with conscious feelings, and we seem unlikely to actually solve that problem before we need rules on companion DMs. But once we produce a very clear surface separations between various DM castes, there may be little pressure to dig deeper. Once DMs dominate the world, and humans fade into the background, DMs might change these rules. But entrenched status rankings tend to persist, so DM caste status rankings may remain for a long time to come. Discuss
Read more- AI Research
MarkTechPost
Google Open-Sources Harness2, a Recursive Agent Harnessing Framework
Google Research released Harness2, an Apache-2.0 framework that rewrites an agent's harness at test time. One refinement step lifted GLM-5 by 11.5 points on WorkBuddy-Bench with no weight updates. The post Google Open-Sources Harness2, a Recursive Agent Harnessing Framework appeared first on MarkTechPost.
Read more - AI Research
Dealroom.co
Claude cracks 25-year mystery behind NIST's elliptic curves
Claude cracks 25-year mystery behind NIST's elliptic curves Dealroom
Read more - Machine Learning
MarkTechPost
How to Make Zero-Training Tabular Predictions with NVIDIA Structured Data Models
In this tutorial, we implement Structured Data Models, NVIDIA’s new library that puts several pretrained foundation models for tables behind one PyTorch interface, and use it to make predictions without training. The library wraps TabICLv2 and KumoTabular for single tables and KumoRelational for whole relational databases, and each of them answers by in-context learning: we […] The post How to Make Zero-Training Tabular Predictions with NVIDIA Structured Data Models appeared first on MarkTechPost.
Read more - AI Products
36Kr
DeepSeek新模型Token消耗降超五成
DeepSeek正式推出大模型新版本V4.1 Flash-1017。在国产人工智能大模型普遍处于迭代关键期的当下，该版本率先完成升级，整体性能实现显著跃升。 此前已有公开数据显示其技术指标，此次DeepSeek更首次自主发布帕累托前沿曲线，直观呈现模型在效率与成本维度的综合表现。V4.1 Flash-1017在此曲线中达到行业领先水平：单位任务综合成本仅为Fable、Astra等同类模型的百分之一至二十分之一，性价比优势达四五十倍以上。本次发布的1017版本虽维持原有API价格体系，但在Token使用效率方面完成深度重构——通过任务感知型压缩、动态上下文裁剪及自适应解码策略，显著降低各类型请求的实际Token消耗。实测表明，多数场景下Token用量较前版下降逾50%，相应TCO水平亦将同步迎来新一轮实质性下探。（新浪财经）
Read more - AI & Society
Effective Altruism Forum
A day in the life of a 2026 software engineer
Published on October 18, 2026 5:12 AM GMT I bike to work and get in at 9am. I check my code review dashboard. There are 5 PRs* to review. 2 are automated cleanups by bots, I approve them without reading them. The other 3 PRs are written by more expensive models (mostly Gemini 4), at the request of my colleagues. I open conversations to have my own agents review them, while I skim the code myself. I catch one business logic mistake. The agents catch 5 technical issues (mostly poor error handling, or inefficiency), and 3 issues that are nits I don't care about (I delete these comments). I approve all the PRs, except the one with the business logic mistake. My colleagues tell their agents to respond to my replies, and then push the results. I check my email. There is one design document someone wants me to review. It was written by AI. It is good, but very long. I skim it, and approve it on a broad strokes basis (it is not a critical part of the system, I don't need it to be exactly correct). There are 4 bugs that have been forwarded to me. 3 are automated alerts. I tell my agents to try to investigate them and create a fix. They create plausible fixes in all cases, and I send my own code out for review. I send the important fixes to my peer who reviews code within an hour, the rest to folks who take a day or two. It is now 11:30, time for lunch. I have done more work this morning than I did in a week last year. At lunch with my team, we discuss AI. The topic rotates between 'man Gemini is so silly I dont know how it could make this mistake' and 'Gemini saved more than my salary in compute, when I spent 5 minutes asking it to make one our jobs more efficient'. People repeatedly say that AI is a bubble, and the models are going to hit a wall. They say models won't get better than they are now. We got work access to Gemini 4.0 only 3 weeks ago. After lunch, I conduct an interview for an intern candidate. I tell the candidate I may need them to close their eyes and describe the algorithm (AI assistance is not allowed since Opus 5.5 would demolish the interview). They struggle with basic syntax, and seem not to have coded much recently. I haven't either. At the end of the interview, they ask me how I see AI changing the industry. I decide to be honest and say that I'm scared shitless. I don't see many parts of my job that I'm better at than the frontier models. The candidate laughs, but then says he agrees. After writing up interview feedback, I spend an hour flipping between reviewing PRs, asking agents to fix old bugs in our system, and reading Astral Code Ten. Oh, no, it was 3 hours and its 6pm. My sense of time is all wonk from flipping between contexts constantly. I go home. *I work at Waymo as a Senior Software Engineer (L5). We would call PRs 'CLs' and would generally refer to Gemini as 'Jetski', which is Google's internal version of Antigravity. Other details are largely correct. Discuss
Read more - AI Products
36Kr
Anthropic Hires Legal Experts to Help AI Agents Negotiate Deals
Anthropic is consulting with law firms to tackle a new challenge: how AI agents should negotiate to secure the best deal outcomes for their human users. Two people familiar with Anthropic's thinking said the company is seeking advice from law firms on how lawyers conduct negotiations, with the aim of representing clients' interests while also achieving fair outcomes. A person familiar with Anthropic's project said the company is exploring "negotiation norms" for interactions between models. As intelligent systems continue to evolve, AI agents will represent users in dealing with other agents whose goals conflict with theirs. The person added that the company is also weighing how much agents should seek human feedback during negotiations, and how wary they should be of other AI counterparts while avoiding paranoid, conspiracy-style judgments. Another person close to the company said Anthropic has also considered an approach in which users would sign documents confirming the matters entrusted to the agent and all the rules it must follow. (Sina Finance)
Read more 
AI SafetyCybersecurity News
OpenCode Security Breach Exposes Third-Party API Keys of its Console Users
OpenCode has disclosed a security breach affecting 1.4% of its Console users, warning that third-party API keys saved in affected workspaces may have been exposed. In an October 17 statement, the company linked the incident to a large-scale distillation attack and said it was working with Anthropic, OpenAI, and Google to address it. The affected […] The post OpenCode Security Breach Exposes Third-Party API Keys of its Console Users appeared first on Cyber Security News.
Read more- AI Products
MarkTechPost
Meet Voxel51 3.0: A Unified Multimodal Data Platform for Physical AI
Voxel51 3.0 merges search, curation, annotation and evaluation of multimodal sensor data into one platform for physical AI teams. It adds Voxel51 Agent with 45+ MCP tools, dataset versioning, and a free self-service tier. The post Meet Voxel51 3.0: A Unified Multimodal Data Platform for Physical AI appeared first on MarkTechPost.
Read more