Lectura

The Finnish Licensing and Supervision Agency (LVV) has ordered Google subsidiary Tuike Finland Oy to stop construction on two AI data centers in Muhos and Kajaani.

The order stems from allegations that the projects cleared over 740 acres of forest without completing required environmental impact assessments. A Google spokesperson admitted the company "fell short of our own high standards" and cited ongoing biodiversity plans including tree planting.

Construction activities such as excavation and soil removal are banned, though planning and surveys may continue. This case highlights tightening environmental scrutiny for European AI infrastructure, distinct from previous bottlenecks focused primarily on power supply.

Anthropic announced two new beta features for Claude on October 8: Dashboards and Motion.

Dashboards allows users to connect data sources like BigQuery and Snowflake via natural language prompts to generate auto-updating live dashboards, showing the underlying queries. This feature is currently in beta for paid users.

Motion generates editable animated explainer videos from text, charts, or images, exportable as MP4 files. This feature is in beta for Team and Enterprise plan users.

Additionally, Docs, Slides, and Design have exited beta and are now available across all plans, including free accounts. The company states these tools have been used to create over 45 million documents.

Anthropic released a reference implementation on October 8 for building daily briefings using Claude Managed Agents (beta), addressing common failure modes in scheduled agent automations.

The guide proposes a six-component architecture—Sources, Destination, Agent, Schedule, Memory, and Guardrails—to tackle issues like silent failures and duplicate reporting. A key improvement is the use of dynamic bookmarks to track the last read timestamp per source, replacing fixed time windows like "last 24 hours" to prevent data gaps or repeats caused by run delays.

It also emphasizes credential isolation via Vaults, ensuring real tokens remain outside the sandbox where code executes. For failed sources, agents are instructed to preserve bookmarks and explicitly note unavailability in the briefing, rather than falsely reporting "nothing new."

This is an official engineering blog post providing best practices for developers building automated workflows with the Claude API.

Anthropic launched OSS Scanner on October 8, a free vulnerability scanning service that sends fully model-generated reports directly to open-source maintainers without human review.

Over the prior six months, Anthropic discovered more than 29,000 candidate vulnerabilities but could only manually triage approximately 6,000 due to human capacity limits. On CyberGym, an academic vulnerability-discovery benchmark, LLM detection rates rose from under 20% to over 85% within a year, creating a backlog of high-quality but unverified reports.

Each OSS Scanner report includes a self-contained reproducer, vulnerability explanation, and candidate patch. In internal validation, Anthropic asked expert penetration testers to review 97 critical and high-severity findings across 48 projects; 85 (88%) met the bar for coordinated disclosure. PostgreSQL and wolfSSL maintainers confirmed most reports were valid, though some severity ratings were inflated.

The service is free and open to eligible projects via a GitHub application. Anthropic acknowledges reports may contain false positives and plans to refine the system based on maintainer feedback.

Manus, the developer of the eponymous AI agent, has announced a funding round exceeding $500 million.

The round was led by private equity firm Boyu Capital, with contributions from Tencent Holdings Ltd. and other investors. Bloomberg previously reported that the deal values Manus at approximately $4 billion, roughly double the price Meta Platforms Inc. reportedly offered last December.

Manus, originally launched as Butterfly Effect Co. Ltd. in 2022, gained viral attention with its multi-step task agent in early 2025. Meta’s attempted acquisition was scrapped in April 2026 after Chinese regulators blocked the deal. Since then, Manus has resumed operations as a standalone company, recently releasing Manus 2.0 which claims to reduce token usage by 23.2% while speeding up processing.

Transformers are slow and memory-intensive on long sequences because the time and memory complexity of self-attention grows quadratically with sequence length. Approximate attention methods attempt to trade model quality for computational savings, but often fail to deliver actual wall-clock speedups.

The paper proposes a missing principle: making attention algorithms IO-aware, i.e., considering reads and writes between different levels of GPU memory. FlashAttention uses tiling to reduce the number of read/write operations between GPU High Bandwidth Memory (HBM) and on-chip SRAM.

It claims to be an exact attention algorithm, requiring no approximation, with results consistent with standard algorithms. It also extends to an approximate version that is faster than all existing approximate methods.

It achieves 3x speedup on GPT-2 (sequence length 1K) and 2.4x speedup on Long Range Arena (sequence lengths 1K-4K). GPT-2 perplexity improves by 0.7, and long-document classification accuracy increases by 6.4 percentage points. It is the first method to enable Transformers to exceed random baseline performance on Path-X (16K sequence, 61.4%) and Path-256 (64K sequence, 63.1%).

Experiments were primarily conducted on GPUs, using specific models and sequence lengths. Beyond certain lengths, some approximate methods may outperform it. One cannot conclude from this that approximate methods are universally ineffective, nor extrapolate these findings to non-GPU hardware or untested configurations.

Suppose you are training a long-context model and notice low GPU utilization but saturated memory bandwidth—this is a typical manifestation of an IO bottleneck. In such cases, reducing data movement between HBM and SRAM is more effective than increasing compute power.

"FlashAttention: Fast and Memory-Efficient Exact Attention with IO-Awareness," Tri Dao et al., NeurIPS 2022. Suitable for readers who want to understand why long-sequence Transformers are slow and how to accelerate them without sacrificing accuracy. Start with the abstract and the IO analysis section.

In his letter to shareholders dated February 26, 1982, reviewing the fiscal year 1981, Warren Buffett posed a highly practical question for judgment: Are reported corporate profits and the economic value actually received by shareholders the same thing? The value of this letter lies in separating accounting presentation, capital usage, and final returns for discussion.

The first concept is that one cannot look only at profits entering the financial statements. Berkshire holds interests in some companies it does not control; the retained earnings of these investee companies may not be reflected in Berkshire's currently reported income, yet they can still translate into shareholder value through the development of those invested businesses. Therefore, understanding a holding requires continuing to ask where the money earned by the investee company went, rather than stopping at how much dividend was received.

The second concept is that retaining profits itself does not equal creating value. The original letter points out that if profits are used inefficiently, market recognition of retained earnings may be low or even negative; if additional capital earns high returns, the value created may exceed the amount retained. The key lies in the use and output of the next dollar; one cannot equate scale expansion directly with successful capital allocation.

This line of thinking applies today when reading a company's reinvestment plans: first distinguish between changes in financial statements and changes in economic reality, then judge the productivity of new capital. It will not automatically provide an answer on whether something is cheap or expensive, nor can this historical letter replace verification against the latest operational data of the target company.

For example, suppose two companies both choose to pay lower dividends and retain more earnings. One puts the money into projects with clear demand and reliable collections; the other merely continues expanding businesses lacking economic viability. Their retention actions are identical, but the outcomes for shareholder value could be opposite. This scenario illustrates only the logic of judgment and does not claim that any real-world company currently fits it.

Also note the time lag. The letter explains that the timing for retained earnings to convert into realized or unrealized capital gains is very irregular, and market recognition varies unevenly across different companies. This means that judging the quality of reinvestment and predicting short-term stock prices require different evidence. This article is written based on excerpts from publicly available letters on the official website.

Recommended reading: "Berkshire Hathaway 1981 Shareholder Letter," authored by Warren Buffett, dated February 26, 1982. Readers studying corporate capital allocation should first read the sections on non-controlled holdings' profits not included in reported income and the value of retained earnings, then read the full text carrying the questions: "How is the money being used, and when can value be verified?"