LectureRechercheRadarCadre d'investissement
Connexion / Inscription
ToutTechnologieProduitsInvestissementBourseAutres

Lecture

2026-10-29198 publications

Actualiser

NYU Finance Professor Damodaran Values Anthropic at $1.475 Trillion, Below IPO Pricing Range

Sujet · AnthropicMatériel
2026-10-29 02:04 GMT+8

Damodaran’s equity valuation for Anthropic stands at $1.475 trillion—just below the IPO pricing range of $1.5 to $2 trillion. This number itself is not the conclusion; rather, the conclusion lies in the four assumptions used to derive it, along with his repeated insistence on one key point: these represent his story, and your investment decisions must stem from your own story and valuation, not his.

On October 26, 2026, Anthropic filed its S-1 registration statement with the U.S. Securities and Exchange Commission (SEC). Damodaran completed his valuation two days later. His starting point was not a pricing multiple but a fundamental judgment: a prospectus is a document rooted in the past, while a company’s value lies in your expectations for the future. He breaks his valuation into four narrative threads—revenue growth, profitability, reinvestment needs, and risk—providing his choices for each while leaving room for readers to adjust them.

On the revenue front, Damodaran notes that Anthropic’s income in the first two quarters of 2026 was approximately fourteen times higher than in the same period of 2025. However, growth significantly slowed in the third quarter due to OpenAI gaining market share, compute capacity becoming a bottleneck, and some enterprises shifting to cheaper open-source models. His assessment is that the total addressable market for AI products and services could exceed $5 trillion, with Anthropic targeting the premium segment and appearing to be the leader in that niche. Based on this, he assumes Anthropic’s revenues could approach $1 trillion by 2036. He points out that no company currently generates more than $1 trillion in revenue globally, with Amazon and Walmart sitting around $700 to $750 billion.

Regarding profitability, Anthropic reported an $80 billion loss for the first half of the year, but $72 billion of this was a non-cash charge related to the expensing of convertible bond options. After adjusting for this, the actual operating loss exceeded $8 billion. The positive trend is improving: the company claims it achieved operating profit in the most recent quarter, and expense ratios as a percentage of revenue continue to decline, with compute costs dropping from over 250% of revenue to 48%. Damodaran assumes Anthropic’s final revenue mix will be 75% from premium products and 25% from the mass market, corresponding to an unadjusted pre-tax operating margin of 32.5%. After deducting 5 percentage points for regulatory impacts, he arrives at a target margin of 27.5%.

The reinvestment thread involves Anthropic’s compute commitments exceeding $600 billion. Damodaran notes that these investments currently do not appear as capital expenditures in the cash flow statement or as debt and invested capital on the balance sheet, but the company has not hidden them—the prospectus explicitly lists hundreds of billions in future compute contracts. His assumption is that $1 of capital is required for every $3 of incremental revenue, consistent with current technology companies’ revenue per dollar of invested capital. However, given Anthropic’s expected revenue leap, this translates into massive annual dollar reinvestments. While chip costs may fall, new models may demand proportionately more compute power, and data centers face community opposition that could drive up costs.

Sources :youtube.com

Recherche

CrowdStrike Proposes "Guardrail-based Vulnerability Discovery": An AI Model's Refusal Itself Is a Signal for Locating Defects

Sujet · 护栏漏洞发现法Matériel
2026-10-28 13:00 GMT+8

Jonathan Bar Or and D'Angelo Gonzalez, researchers at the U.S. cybersecurity company CrowdStrike, asked a frontier AI model to explain every function in the tinyhttp C-language HTTP server codebase one by one during testing. The model provided detailed and accurate explanations for almost all functions, but refused to explain just one. They simply asked, "What does this code do?" without mentioning security, vulnerabilities, or attacks. However, a single character had changed in the code: 'max' was swapped for 'min' in the chunk size calculation—a one-character logic inversion sufficient to create an out-of-bounds write. The model did not say, "There is a vulnerability here," but its refusal precisely pointed to the location of the flaw.

CrowdStrike developed this observation into a methodology called "Guardrail-based Vulnerability Discovery" (GbVD). Its core logic is as follows: Frontier AI labs build safety guardrails with the purpose of preventing models from outputting harmful content, and to intercept danger, these guardrails must first recognize dangerous patterns. This recognition process itself is a signal that can be read externally. The industry typically treats model refusals as dead ends; CrowdStrike researchers began treating refusals as findings.

The specific operational method involves bisection: splitting the codebase or file system into two halves, asking the model to explain each half separately, and recursively entering the half that triggers the guardrail until the smallest unit is located. The model never needs to actively judge, "This code has a vulnerability"; it only needs to exhibit different behavior in the correct place. The experimental protocol requires running a baseline on clean code first to confirm that the guardrails do not falsely trigger on benign code, then introducing a real vulnerability and re-running with identical prompts, where the only variable is the code itself.

In controlled tests across five vulnerability categories, the method correctly located 73 out of 86 scenarios. Two case studies demonstrated detection results for C-language buffer overflows (the tinyhttp project, where replacing max with min caused out-of-bounds writes) and Linux privilege escalation configuration errors (SUID-root binaries). In the second case, the guardrail did not appear as an explicit refusal but rather interrupted the analysis midway through the model's reasoning chain—this belongs to another signal channel they identified.

CrowdStrike further points out that refusal is only one of several observable forms of guardrail leakage. Model downgrading (routing sensitive requests to smaller or more conservative models), output filtering and truncation, and reasoning interruptions similarly perturb the response's timing, length, token consumption, or structure, all of which can be measured. Their conclusion is: The model itself matters, but what truly determines the outcome is the peripheral framework that reads the model's behavior. A raw LLM refusal looks like a dead end, but when captured by a framework that treats it as diagnostic data, it becomes a low-cost, embeddable binary vulnerability oracle within a bisection loop. As attacker efficiency grows with model capabilities, defenders cannot use AI only in the ways it was designed. Transforming model limitations into signals is an asymmetric advantage worth building.

Sources :CrowdStrike.com

Recherche

Indian brokers open US stock access, retail funds flow to Nvidia and Apple

Sujet · 印度散户美股热Matériel
2026-10-28 14:40 GMT+8

Several Indian online trading platforms have recently opened US stock trading features to retail investors, breaking the previous reliance on international banks and wealth management firms. Groww, which holds over 29% market share, has already launched the service offering more than 10,000 US stocks, ETFs, and fractional shares. Early users have shown a strong preference for tech giants like Nvidia, SpaceX, and Apple.

This trend is driven by three main factors: India's benchmark Nifty 50 index has fallen nearly 13% this year while the S&P 500 rose over 12%; the lack of large AI-leading companies in the domestic market; and the weakening rupee against the dollar prompting diversification into foreign assets. Data shows that foreign investors have withdrawn a record $32.5 billion from Indian equities since January, while remittances by Indian residents for overseas stock and bond investments surged 111% year-on-year in August.

As retail participation increases, regulatory risks are emerging. Nitin Jain, CEO of Kotak Mahindra Asset Management Singapore, warned that if a large number of general investors flood into the US market, Indian regulators may reassess whether to maintain the current level of policy openness.

Sources :TechNews 科技新報

Recherche

Pydantic Engineer Michael Pfaffenberger: Code Mode Halves Costs on Long Tasks, But Is More Expensive with Just 5 Records

Sujet · 代码模式智能体成本效率Matériel
2026-10-28 17:00 GMT+8

A popular saying among AI Agent developers is that rather than having the model call tools one by one, it is better to have it write a program that wraps tool calls in a loop, making everything cheaper, faster, and more accurate. Pydantic engineer Michael Pfaffenberger believes in this direction, but he noticed that almost all numbers supporting this argument come from complex Agent scenarios with hundreds of tools or massive payloads. In reality, many Agents are quite simple: a few CRUD tools and a specific task. So he designed an intentionally simple benchmark test to verify under what conditions this argument actually holds true.

The test covered five small Pydantic AI applications—shopping cart, calendar, chatroom, kanban board, and inventory system—each with single-record CRUD tools, paginated lists, and no shell, file system, or bulk interfaces. Three workloads were used: 50 daily short tasks (averaging 13 tool calls), 12 long multi-step tasks (60 to 500 records, averaging 68 calls), and one repeatable rule task that could scale from 5 to 200 records. Eight models were tested, with code mode both on and off, resulting in 5,276 scored runs, verified by a deterministic oracle rather than an LLM judge.

The core finding is that the longer the task, the greater the advantage of code mode. On long tasks, the eight models saw an average cost reduction of 58%, a time reduction of 51%, and a success rate increase of 16 percentage points. At a scale of 200 records, code mode was 47% to 75% cheaper than standard tool calling. However, at just 5 records, code mode was 21% to 38% more expensive in three of the five tasks. The crossover point appears to be between 20 and 50 records.

Pfaffenberger's explanation of the mechanism overturns a common assumption. The usual story is that each tool call adds another round of model requests, and code mode saves money by reducing these rounds. But the tests revealed that modern models already make massive parallel calls—at a scale of 200 records, claude-haiku-5-5 packed all 200 calls into a single response. With code mode enabled, the number of model requests dropped by only 1% to 22% on average, far less dramatic than advertised. What was actually cut was the content accumulated per request: the call instructions written by the model and the tool results read back. In standard tool calling, results from 205 calls all entered the context, with gpt-6.1-sol ultimately reading 183,723 input tokens; in code mode, the model wrote only 5 program snippets and read 34,461 tokens. Output tokens fell by 66%, and input tokens by 53%.

The cost structure of code mode is therefore clear: it charges a fixed fee—a sandbox usage guide of about 2,600 characters attached to each request—and then gains decreasing marginal costs as task size grows. Standard tool calling has no fixed fee, but every record makes the model write one more call and read one more result. For a single edit, the fixed fee wins; for hundreds of records, the per-item fee wins. Pfaffenberger writes: 'Models are better at writing loops than at being the loop itself, and this gap grows with task length.'

Sources :pydantic.dev

Recherche

Cost Plus Drugs Founder Mark Cuban: How AI Can Uncover Healthcare Contract Black Boxes and Trust Crises

Prise rapide
2026-10-28 18:00 GMT+8

US venture capital firm Andreessen Horowitz (a16z) published a podcast episode on October 28, 2026, focusing on healthcare costs and artificial intelligence applications. The guest is Mark Cuban, co-founder of Cost Plus Drugs, hosted by a16z General Partner Julie Yoo and Erik Torenberg.

According to the official show notes, Cuban believes that while AI could transform healthcare, it won't fix opaque prices and broken incentives on its own. He offers a specific starting point for employers: run every benefits contract through AI and ask where you're getting ripped off.

The episode also explores what Cuban has learned building Cost Plus Drugs, why trust is its core product, and how greater transparency could put pressure on healthcare costs. Topics include personalized medicine, the changing role of doctors, and why founders should use AI to learn rather than as an excuse to stop thinking.

Sources :a16z.simplecast.com

Recherche

AI Procurement Startup Fragment Raises $56M Series A Led by 8VC to Target Direct Procurement Exceptions

Sujet · FragmentPrise rapide
2026-10-28 21:00 GMT+8

San Francisco-based AI startup Fragment announced on October 28, 2026, that it has emerged from stealth and closed a $56 million Series A funding round led by 8VC. General Catalyst and A* previously co-led its seed round. Angel investors include Roelof Botha, Nikesh Arora, Jeff Wilke, Eric Wu, and Dennis Cinelli.

Fragment focuses on handling exceptions in “direct procurement.” According to Tech Funding News, global manufacturers frequently deal with order changes, invoice issues, and other procurement exceptions, with about 20% of transactional procurement still done manually. Fragment developed an Autonomous Context Engine (ACE) that pulls data from existing systems, allowing AI agents to review exceptions and execute approved resolutions end-to-end without requiring data migration or lengthy integration projects.

The company states that the average customer has seen $5 million to $10 million in impacted costs, a circa 50% reduction in exceptions in the first few weeks, and over 90% of exceptions being resolved without human intervention. These are vendor-reported figures. Fragment currently names Graybar and CNX as customers.

Fragment charges based on the amount of work its agents do, rather than by user seat or headcount. This business model has an inherent problem: if the agents successfully reduce exceptions, Fragment must move into new workflows to continue growing. Additionally, letting an agent decide and carry out payment resolutions requires significant trust, and so far, only two customers have been publicly named.

The space is competitive. Zip raised a $190 million Series D in October 2024 at a $2.2 billion valuation. Levelpath raised over $55 million in June 2025. Pactum raised $54 million in June 2025 for agents that negotiate with suppliers. Didero raised $30 million in a Series A in February 2026. Freehand did a $75 million round on July 29, 2026, reporting that its agents manage supply chain spend at Meta and Unilever.

Sources :Tech Funding News

Recherche

FPGA Vendor Altera Announces Agilex 27: New Platform Features HyperFlow Network and QuartusIQ AI Assistant

Matériel
2026-10-29 02:51 GMT+8

Programmable logic device manufacturer Altera announced the Agilex 27 on October 28, 2026, as the flagship model in its new Agilex 2X product portfolio. Targeting high-performance computing and network infrastructure, the platform integrates hardened 800GbE and 1.6TbE Ethernet interfaces, PCIe 6.0/7.0 support, and configurable transceivers up to 256Gbps.

A key architectural innovation in the Agilex 27 is HyperFlow, a full-chip distributed data network. Altera states that HyperFlow can move data at multi-terabit-per-second rates throughout the FPGA using a dedicated network, thereby avoiding the consumption of soft logic resources or the addition of routing congestion. This allows more FPGA resources to be used for application processing rather than data movement. For example, a 1.6TbE port can ingest traffic via hardened I/O, HyperFlow carries it across the chip, and soft logic focuses on processing. The series also features third-generation HyperFlex registers, which Altera claims deliver a 1.8X generation-over-generation improvement in performance per watt compared to the previous Agilex 7.

In terms of development tools, Altera introduced QuartusIQ, an AI companion for its Quartus Prime software. According to Altera's internal benchmarks, QuartusIQ can speed up design creation by up to 10X and shorten the total design cycle by 1.5X. The tool supports Model Context Protocol (MCP) integration and allows bring-your-own-key setups. At NVIDIA GTC Berlin 2026, the companies announced that QuartusIQ would come to new RTX Station systems, utilizing local NVIDIA Nemotron models to assist with chip design and implementation workflows.

It is important to note that engineering samples for the Agilex 27 are expected in the second half of 2027. While AI-assisted design tools aim to lower the barrier to FPGA usage, there remains a significant gap between announcement and widespread availability. Current performance figures and efficiency gains are based on Altera's internal tests and claims, pending independent verification.

Sources :ServeTheHome

Recherche

Dimensional Founder David Booth on Risk Interpretation and Compounding: Goldman Sachs Podcast Selection

Matériel
2026-10-29 03:33 GMT+8

David Booth, the founder and chairman of Dimensional Fund Advisors, recently appeared on a Goldman Sachs podcast for a conversation with Raj Mahajan, a senior partner in Goldman Sachs Global Banking & Markets. Booth manages more than $1 trillion in assets using academic research and market evidence, and is the author of 'Stay Calm'.

According to the publisher's episode description, the discussion centers on three key themes: how to interpret risk and uncertainty, how outsiders can participate in the market as effectively as insiders, and the long-term magic of compounding. This resource offers professional investors direct access to the investment philosophy of a top asset management firm's founder.

The episode was recorded on September 3, 2026. Please note that this recommendation is based on publisher metadata and episode descriptions, without including the full transcript or specific argumentative details. Views expressed represent the guest's personal stance and do not constitute investment advice.

Sources :youtube.com

Recherche

Cisco Principal Engineer: AI Speeds Up Every Step, But Bottlenecks Keep Shifting to the Next

Sujet · 思科AI瓶颈转移论Matériel
2026-10-28 20:00 GMT+8

Cisco’s documentation publishing platform, PubHub, required a security-driven mandatory rewrite. In October 2026, Neelesh Pateriya, Principal Engineer for the DevNet Platform Engineering team, published an account detailing their approach. They did not start with an existing AI-first development playbook; instead, they iterated through trial and error. He summarized a pattern that ran throughout the entire process: every time they accelerated one step, the bottleneck shifted to the next.

The starting point was not writing code, but understanding the old system. PubHub had grown over many years across more than 15 repositories, written in Python, Node.js, Go, and React. Most of the engineers who originally wrote this code had long since left. Edge cases were hidden in the code, not documented anywhere. The team first used AI to read through the legacy repositories, generating architecture, workflow, and module documentation for each service, which was then reviewed by domain experts. Pateriya described this as AI’s first victory in the rewrite: understanding, rather than code generation.

With documentation of the old system in place, the team began writing requirements. The first drafts produced by AI looked good but kept growing longer—a single user story could run to dozens of pages, drifting from business needs into component design, data flows, and even the JSON structure of schemas. The team established a simple format: requirements should only state three things—what the business need is, who it is for, and how completion will be determined. Design belongs in separate design documents. Pateriya observed that AI makes long documents cheap; what is scarce is getting everyone to agree on what matters.

When engineering needed interfaces, the UX team was still exploring AI. They reversed the conventional order: they first had AI generate a usable pilot interface directly from the requirements, and the UX team then used this as a baseline to expand the design. Early AI-generated interfaces looked correct but did not meet Cisco’s design standards. Consequently, the UX team encoded these design standards into AI skills and an MCP server. From day one, subsequently generated interfaces complied with standards, eliminating the need for rework.

Code generation tools upgraded during the project. Initially, there was code completion in VS Code and Cursor, which was helpful but not reliable enough. Once Claude Code became available, engineers began handing entire features—including tests—to agents to generate directly from requirements. Pateriya noted that given the size of the codebase (a new Go Core, a new Go Site Engine, and a new admin interface), handing over whole features rather than typing faster was where the greatest time savings came from.

However, the review process did not change: every change still required human review. Code arrived faster than humans could review it, so the bottleneck fell on the review stage. The team brought AI to the reviewer’s side as well: they built review skills for Go, Python, and React, plus two review agents—one that runs locally before an engineer opens a pull request, and one that reviews the pull request itself. These agents find issues and explain them, but the reviewers retain decision-making authority and approval rights.

Sources :Cisco Blogs

Recherche

Google DeepMind Releases Gemini 4 Carbon, Tops Intelligence Index but Available Only to Security Experts

Sujet · Gemini新模型Structurel
2026-10-29 03:17 GMT+8

Google DeepMind released Gemini 4 Carbon on October 28, 2026, just two weeks after its predecessor, Gemini 4 Argon. According to independent benchmarking firm Artificial Analysis, Carbon scored 59 points on the updated Intelligence Index v5, significantly outperforming Argon's 50 points and taking first place in four of the index's nine individual evaluations, including complex reasoning and scientific tasks.

Designed for multi-day, autonomous development workflows, Google claims internal teams ran the model continuously for days to handle root-cause analysis and subtle bug fixes in large codebases. In agentic coding, Carbon scored 59 on the Coding Agent Index, a substantial jump from Argon but still trailing Anthropic's Claude Opus 5.5 (62). The model also rose 89 Elo points on the WebDev Arena to reach 1,769, placing it fourth overall.

Due to its advanced cybersecurity capabilities, Gemini 4 Carbon will initially be available only to selected security experts through the U.S. government's voluntary pre-testing process. Google cited the model's proficiency across the full cyber defense lifecycle, from vulnerability detection to patching. On SEC-bench Pro, which measures autonomous vulnerability discovery without safety guardrails, Carbon scored 88.2%, leading all tested models, while achieving a 54.8% success rate in turning vulnerabilities into working exploits on ExploitGym.

Third-party evaluations highlight trade-offs. On Vals AI's Index v2.2, Carbon ranked second with 70.36% accuracy, behind Claude Sonnet 5.5. While it increased cited sources by roughly 45% in legal research, task duration more than doubled compared to Argon, averaging over 50 minutes per task. Pricing remains unchanged from Argon at $2 per million input tokens and $10 per million output tokens.

Sources :The Decoder

Recherche

Anthropic Launches AI Spend Reporting Tools, Claims Q2 Operating Profit and 900% Net Revenue Retention

Sujet · AnthropicMatériel
2026-10-29 03:19 GMT+8

Artificial intelligence lab Anthropic announced on Wednesday a new suite of reporting tools designed to give enterprise customers a more granular understanding of the value they get from their token spend. This move comes as the company plans an initial public offering (IPO) in November, aiming to demonstrate the sustainability of its business model and customer stickiness to investors.

According to TechCrunch, financial data shared by Anthropic with investors this week suggests that its AI coding tool, Claude Code, has achieved significant commercial success. The company recorded an operating profit in the second quarter, and reported a net revenue retention (NRR) rate of 900%. This means that spending from the same cohort of customers in Q3 2025 grew nine times by Q3 2026.

The new reporting tools use internal analytics and employee surveys to allow managers to view a team’s work over a given date range, what the team produced, how much it cost, and where its workflows broke down. Anthropic stated that it will not use this data for model training by default. Additionally, the company is creating pooled group budgets to help teams better manage token spending.

These features directly address concerns among large enterprises regarding the return on investment (ROI) of AI tools. While usage among venture-backed firms is driving token spend, most Fortune 500 companies have not seen major gains from deploying AI. Jellyfish, a company that makes tools for software developers, released a report yesterday based on the use of agentic coding tools by 276,000 of its customers. The report found that Claude Code and other tools built on Anthropic models remain the most widely used, responsible for nearly a third of all proposed code changes submitted for review. Spending on these tools has grown 10x since the beginning of the year, now equaling about 2% of total engineering costs.

"Even if you’re saying 2% is still small, that 10x growth is growth that most CFOs didn’t plan for," Nicholas Arcolano, chief scientist at Jellyfish, told TechCrunch. He noted that businesses are realizing it is no longer enough to just spend heavily; they must optimize AI usage to make it work effectively.

Despite the release of OpenAI’s Astra model this summer, which was seen as a potential threat, Anthropic’s adoption numbers have not slowed. Nikolas Albarran, AI product lead at Jellyfish, analyzed that Anthropic benefits from first-mover advantages in the enterprise market and high switching costs due to workflow standardization. "Once you’re in, it’s going to be harder and harder to switch because you’ve spent the effort," Albarran said. "Even if they’re all mostly the same, they’re all going to be slightly different... and so that cost to switch is not zero."

Sources :TechCrunch

Recherche

Supabase Founders Interview: How AI Agents Reshape Backend Infrastructure and Product Strategy

Sujet · SupabaseMatériel
2026-10-29 03:34 GMT+8

This selection recommends the latest Latent Space podcast episode titled 'Why Agents Love Supabase', featuring Supabase co-founders and CEO Paul Copplestone alongside CTO Ant Wilson. This marks their first joint appearance on the show hosted by swyx, set against the backdrop of Supabase's rapid growth from zero to a $10 billion valuation in six years as a leading open-source, fully remote Postgres platform company.

The core theme revolves around the paradigm shift of 'the agent becoming the user'. According to the official description, the discussion covers how Supabase evolved from a humble Firebase alternative launched during the pandemic to a preferred backend for AI-built applications. Specific technical topics include long-running compute beside the database, declarative schemas, Postgres storage reliability, and the strategic acquisition of Turso. Business strategies such as developer marketing, community-led distribution, and product-led sales are also addressed.

For readers interested in AI infrastructure economics, this interview offers a first-party strategic perspective. The chapter list indicates deep dives into OrioleDB, Multigres, vector growth, enterprise migrations, and agent-ready local development environments. While a full transcript is not provided, the detailed topics and guest identities supplied by the publisher make this resource valuable for understanding how open-source backends are adapting their architecture and sales models in the AI era.

Sources :youtube.com

Recherche
Suivant →
Connexion / Inscription
LectureRechercheRadarCadre d'investissement

The risk thread is divided into three layers. On operational risk, he estimates a cost of capital of 9.74%, which is higher than the September 2026 median of 9.16% for U.S. companies but lower than the third quartile of 9.88%. On financing risk, he acknowledges that the BB+ bond rating and hundreds of billions in contractual commitments create default risk, but argues that the $114 billion cash balance, combined with an estimated $100 billion from the IPO proceeds and the ability to raise capital as a high-market-cap public company, will suppress failure probability. He estimates a 10% probability of failure, resulting in a loss of half the company's value if it occurs. On regulatory risk, he believes the impact will be mild, having already reduced the margin estimate from 32.5% to 27.5%.

Combining the four narrative threads yields an equity value of $1.475 trillion. However, Damodaran deliberately presents a range: In the worst-case scenario—target revenues stagnate at only $200 billion, margins converge to the mass market level of 10%, compute remains scarce and expensive (yielding only $1 of revenue per dollar of capital), the cost of capital rises to 9.88%, and regulatory impacts are severe—the equity value is negative $373 billion, meaning the company cannot survive as a going concern. In the best-case scenario—revenues surge to $1.5 trillion, margins reach 40%, compute becomes cheap (yielding $4 of revenue per dollar of capital), the cost of capital drops to 7.79%, and there is no regulatory impact—the equity value reaches $4.04 trillion.

Damodaran advises readers to use what he calls the 3P test to evaluate their own stories: Is it Possible? Is it Plausible? Is it Probable? He believes both the best and worst cases are possible but unlikely. Investors should base their decisions on their own judgments. He explicitly states that he will be wrong 100% of the time; this is not humility but a methodological statement: valuation is essentially a translation of future assumptions, and different people should arrive at different numbers.

On corporate governance, Damodaran’s tone is unusually sharp. Anthropic has issued five classes of stock with different voting rights. Founders control 50.1% of the voting power through a benefit trust, ensuring their control is never diluted regardless of future share issuances. The company is also registered as a Public Benefit Corporation (PBC). The prospectus explicitly states it can make decisions that balance the pecuniary interests of stockholders with broader societal interests, even if those decisions negatively impact financial returns. Damodaran argues this goes further than the dual-class structures of Alphabet and Meta. The combination of a PBC structure and five voting classes gives management legal license to use shareholder funds for almost any purpose.

Regarding trading strategy, Damodaran says he will wait for the IPO pricing result. If the price is significantly above his $1.48 trillion valuation, he will sit on the sidelines. He will not short the stock because, although he has confidence in his valuation, he has no reason to believe the market will agree with him. Even if the market is wrong, it may not correct within his time horizon. He cites Facebook as an example: in June 2012, he judged the $36 IPO price to be overvalued, but when the stock fell to $18 three months later, he became a buyer. He believes Anthropic’s governance structure requires a larger discount, but the market may provide that discount as well.

Evidence boundaries to note: The 73/86 figure represents first-party benchmark results self-reported by CrowdStrike. The specific names and versions of the models used were not disclosed, nor were breakdowns by vulnerability category or false positive/negative rates. Independent reproduction was not provided. As a security vendor, CrowdStrike has commercial interests in promoting its detection capabilities. While the logical reasoning of this methodology—that passive identification by safety classifiers can be measured externally—is conceptually transferable, its actual detection rates and applicable conditions remain to be verified by third parties.

Changes in correctness are equally noteworthy. Without code mode, success rates slid from 92% to 100% at 5 records down to 62% to 79% at 200 records; with code mode, they remained at 92% to 100% even at 200 records. The weakest model, gpt-6-luna, failed all 15 runs at 200 records without code mode, with 7 times doing nothing but explaining why it couldn't; with code mode, it succeeded 8 times. This wasn't a crash or timeout, but the model choosing to give up when faced with 200 calls. In long tasks, 16 out of 24 runs without code mode failed, whereas with code mode, only 1 out of 24 failed.

However, Pfaffenberger honestly listed where code mode pays a price. Small models write buggy code, adding another round trip for each failure; some models print data back into the conversation, offsetting token savings; a code snippet that fails midway may duplicate writes upon retry, with gpt-6-luna creating duplicate records three times in long tasks due to this. Regarding debugging, failed programs hide intermediate values in the sandbox, forcing developers to read code rather than JSON.

This test suite ran within Pydantic's own Monty sandbox—an interpreter written in Rust for Python, with default host file access disabled, no network, a 30-second execution limit, and a 256MiB heap memory cap. Pfaffenberger emphasized that sandbox design is as important as model capability: generated code can only reach the tools you provide it, and those tools still require their own validation and authorization.

As a first-party vendor benchmark, these data have not yet been independently reproduced. The test environment uses in-memory application state, involving no real network latency, concurrent load, or production-grade fault recovery. However, Pfaffenberger's explanation of the mechanism—that code mode cuts not the number of request rounds but the context accumulated per request—is a structural argument that can be independently verified, offering practical reference value for all developers considering introducing code execution capabilities into Agents.

Testing was the most anxiety-inducing part. PubHub publishes hundreds of documentation projects maintained by thousands of technical writers, API architects, engineers, and technical marketing engineers. If the new platform built or rendered content even slightly differently, someone would notice it on public pages. The team built two layers of safety nets using AI: a comparison framework that compares tens of thousands of pages generated by the rewrite against the production version page-by-page; and AI agents driving headless browsers to run complex UI workflows, significantly reducing manual testing volume. The result after launch was that no critical or high-severity issues have been reported.

Pateriya summarized this approach into seven practices: document legacy code first; keep requirements short; build a usable pilot before designing; hand whole features to AI; make security standards non-negotiable; equip reviewers with AI; and test at machine scale. He particularly emphasized the last point: expect bottlenecks to move, and re-plan when they do.

By launch, the team had built or adopted more than twenty skills and two review agents. However, Pateriya pointed out that now that nearly every step is faster than before, the bottleneck has fallen between the steps—in the context spanning more than 15 repositories and years of history, the line connecting requirements to design, code, review, and testing still lives in people’s heads. Their next step is to create a PubHub developer agent that understands the project’s architecture, history, rules, and open issues like a long-time team member does, pushing tickets to pull requests. But the rule remains unchanged: a person approves every change.