Reading

In a documentary released by Machine Learning Street Talk on October 26, 2026, host Tim Scarfe converses with Kevin Hartnett, author of The Proof in the Code, exploring the intersection of formal verification and artificial intelligence in mathematics. Hartnett, a senior writer at Quanta Magazine, focuses his new book on the history and current state of computer-assisted proofs.

The film traces the development of the Lean theorem prover, from its origins as a side project at Microsoft Research to its status as a foundational tool mathematicians now use to check their own field. Through perspectives from participants including Leonardo de Moura (Lean's creator), Jeremy Avigad, Kevin Buzzard, and Paul Lessard, the program showcases the construction of Mathlib, Peter Scholze's Liquid Tensor experiment, and Terence Tao's Equational Theories project.

The second half turns to AI's impact on mathematics, specifically AlphaProof and autoformalisation techniques. The program introduces a core concept: the 'formalisation gap,' defined as the distance between the mathematics people mean and the statement a checker certifies. This framework aims to explain the true meaning of verification when machines generate proofs that no human can read.

For readers interested in the boundaries of AI reasoning capabilities and evaluation credibility, this resource provides narrative context for understanding current technical limitations. While a full transcript is not provided, the official description clarifies the thematic arc, making it suitable as an introductory material for grasping the challenges of AI mathematical verification.

Prof. Riccardo Rebonato, currently Professor of Finance at EDHEC Business School, delivered the keynote address at the EPAT Online Convocation 2026 hosted by QuantInsti. The event celebrated graduates of Batches 63 to 68. Rebonato’s background is distinctive: he began as a physicist with a doctorate in nuclear engineering before becoming part of the first generation of quantitative analysts. His career included complex derivatives trading at Barclays, serving as Global Head of Market Risk and Global Head of Quant Research at RBS, and leading Rates and FX Analytics at PIMCO.

In this talk, which he describes as "not a success story," Rebonato reflects on what he wishes he had done differently over 25 years in finance. The session outlines his journey from nuclear physics to trading interest rate derivatives, and later moving into risk management. A significant portion of the discussion addresses how to read a firm's culture, specifically looking at risk tolerance, collaboration, and bonus structures. He also touches on the experience of living through the 2008 financial crisis from inside a bank that required rescue.

For new quants and professionals in quantitative finance, the episode provides specific advice on taking risk management seriously and maintaining control over powerful new tools, including artificial intelligence. The publisher's description highlights chapters on how firms treat risk-taking and mistakes, the interplay between incentives and behavior, and broader questions about whether certain practices are good for the individual, the firm, and society. This resource is recommended for its authoritative perspective on career development and institutional dynamics in quantitative finance.

Samil PwC, the South Korean member firm of the global consultancy, released a report on October 26, 2026, warning local semiconductor companies to defend profitability as Chinese rivals scale up production and strengthen market share. The report noted that Chinese competition would likely heat up first in conventional memory, mature-node manufacturing, and packaging and testing.

To stay ahead, Samil PwC urged South Korean semiconductor companies to sharpen cost competitiveness in commodity memory while doubling down on cutting-edge sectors such as advanced packaging and high-bandwidth memory (HBM), which are used alongside AI processors in a stacked configuration. The report stated that China still lagged in these advanced segments, where customer validation and stable, high-yield mass production remained barriers.

Data showed that Chinese memory makers were already converting technical gains into commercial share. According to Counterpoint Research, ChangXin Memory Technologies (CXMT) captured 10 per cent of global DRAM revenue in the second quarter, more than double the 4 per cent recorded a year earlier, making it the world’s fourth-largest supplier behind Samsung, SK Hynix, and Micron Technology. In flash storage, Yangtze Memory Technologies Co (YMTC) ranked third globally by NAND shipment volume in the second quarter, trailing only Samsung and SK Hynix. YMTC’s share of global NAND revenue climbed to 14 per cent in the first quarter of this year, up from 9 per cent a year earlier.

Counterpoint projected that YMTC’s server-related revenue could approach 39 per cent of its total sales by 2028, up from virtually zero in 2024, fuelled by an expansion into storage for AI data centres. MS Hwang, research director at Counterpoint, said YMTC was benefiting from rising prices and stronger domestic order volumes, with enterprise storage presenting a lucrative channel to ride the AI wave. However, he added that winning over international clients beyond China would remain a key test of YMTC’s ability to challenge Samsung on the world stage.

Japanese automaker Honda demonstrated its "Honda Avatar Robot" earlier this month at the CEATEC 2026 conference in Japan. The system combines high-degree-of-freedom dexterous hands with AI-assisted controls, enabling users to perform complex physical tasks remotely. Each hand features 16 degrees of freedom and 16 actuated joints, with speeds up to 180 degrees per second. During demonstrations, the robot performed intricate tasks such as threading a needle, tightening small screws, and manipulating bolts.

This design aims to address a persistent manipulation problem in robotics: while machines are becoming smarter at navigating complex environments, handling delicate objects or intricate tasks remains challenging. George Chowdhury, an analyst at ABI Research, stated that teleoperation is "critical" for the commercial viability of humanoids due to the fragility of current AI models. "When autonomous systems fail, decision makers need the reassurance that a production line will not be down for an extended period," he said.

Abdil Tunca, an analyst at Gartner, noted that Honda's approach could open applications where full autonomy is too expensive or technically impractical, such as aerospace manufacturing, electronics assembly, and pharmaceutical laboratories. While not entirely new, Tunca believes Honda's solution has greater commercial potential. He predicted that fewer than 20 companies will scale humanoid robots into production for manufacturing and supply chain applications by 2028, underscoring the difficulty of moving beyond demos.

The industry is shifting toward "shared control architectures," where AI handles routine movements and humans intervene only during exceptions. Chowdhury cited examples like 1X, which plans to enter home settings this year but acknowledges autonomous AI isn't ready, thus relying heavily on teleoperation to capture training data. Agility Robotics uses a "horse and rider" approach, where humans direct the robot during exceptions without taking direct control. If successful, this shift could allow one operator to control an entire fleet, significantly changing the economics of robotic deployment.

According to ZDNet Korea, StarCloud, a US startup building low-Earth orbit data centers, has completed a $250 million funding round. The round was led by Manhattan West Ventures, with participation from Nvidia, Cisco Investment, Benchmark, and EQT. Doonam Partners, an investment subsidiary of South Korean crypto exchange operator Upbit's parent company Doonam, was the only Korean investor in this round.

Founded in 2024 by engineers from SpaceX Starlink and Airbus, StarCloud aims to address the physical limitations of ground-based data centers, such as power shortages, land constraints, and cooling water scarcity. The company proposes using space environments for 24-hour solar power and radiating waste heat into deep space without coolant, aiming to build the most cost-efficient next-generation compute infrastructure within ten years.

Technically, StarCloud launched its first 60kg satellite, 'StarCloud-1', last November, successfully running an Nvidia H100 GPU in low Earth orbit. It also integrated Google's open model Gemma to analyze real-time radar data from Capella Space, verifying on-orbit inference capabilities for detecting distressed ships and wildfires. The company has applied to the US Federal Communications Commission (FCC) for permission to deploy a constellation of 88,000 satellites with 20GW of orbital compute capacity.

Philip Johnston, CEO of StarCloud, noted that South Korea has the second-highest number of paid ChatGPT users globally, indicating rapid AI adoption and strong demand for compute resources. He expressed hope that Doonam Partners would bridge StarCloud with local partners. Lee Kang-jun, representative of Doonam Partners, emphasized StarCloud's execution capability in attracting talent and capital ahead of big tech firms, stating the investment could open channels for collaboration between global frontier technology and Korean industries.

The U.S. Department of Commerce recently awarded a total of $260 million to two fusion energy startups, Helion Energy and Commonwealth Fusion Systems (CFS), under the CHIPS and Science Act. The funds are designated for developing advanced semiconductors to help power and control their reactors. Helion received $175 million, while CFS secured $85 million.

Unlike previous non-dilutive government grants, these transactions come with new conditions: both companies agreed to provide the U.S. government with non-voting equity stakes in return. According to TechCrunch, Helion stated that the government will receive approximately 1% equity based on its Series G round, which valued the company at $15.5 billion. The size of the government's stake in CFS is unknown, as it is based on the company's most recent valuation, which has not been shared publicly.

This shift marks a change in how the Trump administration supports strategic industries. Historically, government funding was viewed as "free lunch" that did not dilute existing investors' ownership, aiming to boost innovation without directly capturing commercial upside. However, precedents set by MP Materials, USA Rare Earth, and Intel show that equity-for-funding deals are becoming more common. David Kirtley, co-founder and CEO of Helion, noted that this structure aligns the interests of the company and U.S. taxpayers around commercial success.

For investors in the fusion sector, this introduces complexity into capital structures. While the $260 million is small compared to the $4 billion raised by CFS and $3.2 billion in committed capital by Helion, the signal is significant. It indicates renewed active federal engagement but raises new questions about priority rights, dilution effects, and the status of government equity in scenarios like liquidation or subsequent financing rounds. Venture capitalists must now weigh whether accepting funds with equity strings attached is worth the trade-off and how it impacts future exit paths.

Cybersecurity company Surfshark has released a study indicating a rapid global spread of AI-generated election deepfakes, posing new threats to democratic processes and electoral fairness. According to the report, which analyzed media-covered incidents between January 2020 and September 2026, a total of 235 election-related deepfake cases were identified across 46 countries.

The data shows a significant acceleration in 2026. Cases confirmed in the first nine months of 2026 reached 66, exceeding the annual total of 52 cases recorded in 2025 by 27%. Surfshark attributes this surge to the increased accessibility of AI-based video and audio generation tools, which have lowered the technical barrier for creating manipulated content.

Geographically, the United States leads with 87 cases, accounting for 37% of the global total. India follows with 19 cases, while South Korea ranks third with 16 cumulative cases since 2021. In the Asia-Pacific region, 73 cases were confirmed across 15 countries, with incidents concentrated in India, South Korea, Taiwan, and Indonesia.

Social media platforms remain the primary vector for dissemination. Among cases where the distribution path was identified, 84% spread via SNS. Specifically, X (formerly Twitter) accounted for 49% of these instances, followed by TikTok (29%), Facebook (22%), Instagram (12%), and YouTube (7%). Note that percentages may exceed 100% due to multi-platform distribution.

Luis Costa, Head of Research at Surfshark, noted that the democratization of creation technology is a key driver, stating that what once required studio-level resources can now be achieved with an app and minutes. He warned that deepfake content often prioritizes emotional reaction over fact-checking, urging users to verify suspicious candidate videos through reputable news sources before sharing.

The study combines data from the AI Incident Database, Resemble AI, and OECD sources. Surfshark emphasizes that these figures represent a conservative estimate based on publicly reported cases, suggesting the actual scope of unreported deepfake influence may be broader.

US enterprise software giant Oracle announced a series of new AI features at its "Oracle AI World 2026" conference on October 26, 2026, aiming to accelerate the transition from software preparation to value creation for enterprise customers. Oracle CEO Mike Sicilia stated in his keynote that the goal of AI is to reduce the time customers spend configuring software, allowing them to focus more on business outcomes.

Key announcements included the general availability of "Fusion Clo," an agentic application execution environment, and 25 new agentic apps. These applications cover finance, human resources, supply chain, sales, and service, designed to help enterprises handle complex tasks at scale and economically. Fusion Clo provides new runtime support enabling agents to perform deep research, calculation, simulation, modeling, and continuous replanning, thereby executing more specialized workflows.

Additionally, Oracle released the "Oracle Fusion Setup and Configuration Registry" (FSCR), a public GitHub repository. FSCR converts Oracle's expert knowledge and best practices into reusable AI skills, allowing AI agents to directly read, plan, execute, test, and validate the setup and configuration of Oracle Fusion applications. This move addresses the pain points of outdated static documentation and time-consuming manual configuration, enabling system integrators and customers to work on the same up-to-date product foundation.

Chris Leone, Oracle's VP of Application Development, noted that next-generation enterprise AI focuses on letting software handle more tasks while users concentrate on setting goals and policies. Through tools like Fusion Companion, users can interact with Oracle apps similarly to ChatGPT, with the system automatically selecting and coordinating necessary transactions, workflows, and approvals. While these features are now officially released, their actual impact on enterprise operational efficiency remains dependent on subsequent customer adoption and independent evaluation.

US tech giant Oracle officially announced its 'Deep Trust' AI Architecture on October 26, 2026, during its Oracle AI World conference. This blueprint aims to make trust an inherent property of the enterprise technology stack, rather than leaving it as a responsibility for each individual AI agent or application to implement.

As organizations increasingly deploy AI agents and AI-generated applications onto existing architectures, they face growing challenges in protecting sensitive information, enforcing business rules, and ensuring accurate answers. Oracle states that the Deep Trust AI Architecture addresses these issues by embedding trust where it cannot be bypassed or compromised by AI models, deep within four critical layers: software platform, data, data access, and application development.

The architecture integrates existing technologies such as Oracle AI Database, Oracle Autonomous AI Database, and Oracle APEX AI Application Generator. Specifically, Oracle Deep Data Security allows organizations to define end-user access rules directly in the database and enforce them on every SQL request. Additionally, Oracle SQL Assertions enable the enforcement of business rules (such as spending limits) at the database level, regardless of which application, user, or agent submits the request.

In terms of application development, Oracle APEX AI Application Generator enables AI coding agents to translate natural-language descriptions into APEXlang declarative definitions, which are then executed by Oracle APEX's centralized runtime engine. This means AI generates only the application definition, not the running code, thereby reducing the security risks associated with maintaining large volumes of AI-generated execution code.

"AI models can now enable amazing productivity and innovation. However, enterprises can't realize these benefits if they can't trust what AI generates. AI trust is the biggest challenge we face today," said Juan Loaiza, Executive Vice President of AI Database Technologies at Oracle. He added that building trust into the underlying software platform, data, data access, and application development layers enables organizations to put AI to work while maintaining control over information and actions.

Shares of US memory chip maker Micron Technology rose 4.2% to $914.4 in Monday trading, extending a 2.5% gain from Friday. The rebound followed investor reassessment of the company's strategy against Chinese competitor ChangXin Memory Technologies (CXMT) and improved visibility into future supply agreements. Previously, Micron had faced pressure amid broader doubts about artificial intelligence-related investments and fears that CXMT could capture market share through new architectural breakthroughs.

Wedbush Securities analyst Matt Bryson stated that concerns over CXMT's proposed "4F square" (4F2) memory design were likely "severely overblown." Micron has told analysts it rejected the format due to limitations in scaling and density compared with its preferred three-dimensional dynamic random-access memory (DRAM) approach. Bryson argued that if Micron's view is correct, it suggests China is seeking an alternative path to bridge its process disadvantage rather than posing an immediate equivalent threat. Bryson maintains an Outperform rating on Micron with a $1,400 price target, based on a six-times P/E multiple of his fiscal 2028 earnings estimates.

Meanwhile, Bank of America assigned Micron a Buy rating after meetings with senior management in New York and Boston. Discussions revealed that Micron's supply and capacity agreements now extend approximately five years beyond 2030, with negotiations already underway for 2031. These agreements include commitments covering memory bit quantities and cash payments. Micron expects them to account for around 35% of total revenue, with a target of increasing that share to 50% by 2030. Such arrangements offer greater visibility into future demand, although they do not remove the broader risks facing the memory-chip industry.

Additionally, Micron outlined a technical roadmap covering advances in high-bandwidth memory (HBM), DRAM, and NAND. The company stated it can produce HBM stacks ranging from four-high to 16-high configurations. Its next DRAM node, known as 1-delta, is scheduled to enter volume production in the second half of 2027, while its G11 NAND generation is expected to approach 400 layers and deliver twice the bit capacity of its G9 generation by the end of 2027. These technological progressions, combined with long-term supply agreements, provide key points for investors assessing Micron's future growth prospects.

AI agents have no internal memory. Every time they call a tool, read the result, or decide on the next step, they must resend the model their tool definitions, instructions, all previous calls, and results. By turn 20, the model re-reads everything from turns 1 through 19, and you pay for those repeated tokens again and again. The AWS Machine Learning blog uses the movie Groundhog Day to describe this predicament.

Prompt caching saves exactly this redundant computation. When a model reads a prompt, it first performs a write called prefill, processing each input token to build key-value vectors as its working state. Caching stores this state; if subsequent requests start with identical tokens, the service loads the saved state directly, processing only the new additions. Amazon Bedrock offers two modes: implicit, where the service automatically selects the endpoint of reusable prefixes, and explicit, where code marks the endpoint using prompt_cache_breakpoint. These can be used in combination.

The first rule for maintaining hits is prefix stability down to the byte. Tool definitions and instructions should be placed at the beginning, with conversation history and new content appended afterward. Placing a timestamp, username, or request ID inside the cache prefix will invalidate everything that follows. A cacheable prefix must contain at least 1,024 tokens to qualify. The author emphasizes that MCP servers and other tool delivery methods need to be indexed to maintain serialization consistency.

When the tool list changes mid-session, the problem becomes tricky. Adding, removing, or reordering tool definitions invalidates the cached prefix. AWS offers two paths: if known tools are simply invoked in different steps, keep the complete TOOLS list unchanged and use the allowed_tools parameter to restrict the subset the model can actually call; if entirely new tools are needed during the session, append their schemas via additional_tools inputs while keeping the initial TOOLS static. In implicit mode, the cache extends to the new user message (including the new tool definition). However, the additional_tools item itself cannot carry a breakpoint marker.

Sub-agents and side calls are another common scenario. Multiple helpers need to share the same background context but have distinct tasks. The approach is to send a warm-up request to write the shared prefix to the cache, confirm non-zero values for cache_write_tokens or cached_tokens in the usage stats, and then launch helpers in parallel. Each helper reuses the same prompt_cache_key, model, and TOOLS, appending only its own task at the end. If the parent agent requires a summary or evaluation after completion, deep-copying the full history and appending instructions allows the helper to hit the cached historical portion.

On October 13, 2026, a jury at the US Federal District Court in Delaware delivered a verdict finding that UK-based chip IP licensor Arm Holdings violated its Architecture License Agreement (ALA) with Qualcomm and intentionally interfered with Qualcomm's business relationship with its customer Meta. However, the jury awarded only nominal damages: $1 each for the breach of contract and interference claims, totaling $2. The critical unresolved issue is whether Judge Maryellen Noreika will grant Qualcomm's requested remedy of suspending royalty payments to Arm for up to five years.

The lawsuit stems from fundamental disagreements over intellectual property commercialization models between the two semiconductor giants. Arm primarily profits from licensing Instruction Set Architectures (ISAs) through two main modes: Technology License Agreements (TLA), which allow customers to use off-the-shelf Cortex CPU designs, and Architecture License Agreements (ALA), which permit customers to design unique CPU cores based on the underlying architecture while maintaining software compatibility. Qualcomm acquired ALA rights through its purchase of Nuvia and used them to develop its custom Oryon CPUs for Snapdragon smartphones, AI PCs, and automotive platforms. Arm accused Qualcomm of abusing ALA terms to avoid per-product royalties and issued a termination warning in 2024.

For Qualcomm, the favorable jury verdict removes the risk of an injunction against manufacturing or selling Oryon CPUs, clearing legal hurdles for continuing to launch devices with its self-developed chips. In a statement, Qualcomm welcomed the jury's finding that Arm failed to meet its long-term contractual obligations and intentionally disrupted its business, calling it "a win for innovation and competition." Conversely, Arm stated that the jury rejected most of Qualcomm's claims and awarded only $2 in total damages, arguing this validates Arm's good faith in fulfilling contracts and suggesting Qualcomm's lawsuit aimed to deprive Arm of billions in legitimate royalty income.

The real financial threat lies in the pending royalty suspension order. Contractual clauses suggest that if Arm is found to have committed a material breach, Qualcomm may be entitled to halt royalty payments for up to five years. Qualcomm is one of Arm's largest revenue sources, accounting for approximately 10% of Arm's global sales in fiscal year 2024. If the judge approves this remedy, Arm faces a substantial cash flow gap, potentially impacting its R&D funding reserves and market valuation since its September 2023 IPO. Additionally, this protracted litigation has prompted other semiconductor designers and cloud service providers to accelerate investments in open-source ISAs like RISC-V to reduce legal dependency on a single IP supplier.