LeituraPesquisaRadarFramework de investimento
Entrar / Cadastrar
TudoTecnologiaProdutosInvestimentoBolsaOutros

Leitura

2026-11-0585 posts

Atualizar

DevRev CTO Ahmed Bashir Says Engineering Team Has Fully Shifted to Codex and Dropped Cursor

Tópico · DevRevVisão rápida
2026-11-04 08:00 GMT+8

In a customer case study published by OpenAI, DevRev Chief Technology Officer Ahmed Bashir described a specific tool migration process: his engineering team fully shifted to Codex and dropped its previously held Cursor licenses after exploring multiple coding tools. Bashir emphasized that this shift was not mandated by management but was the result of engineers independently comparing options and naturally converging on a choice.

Bashir attributed this process to developers' direct judgment regarding tool quality. He stated, "Developers are really just looking for the best tool, and you can't fool them." He noted that no one in the team raised objections to continuing with Codex, believing it offered "far and away the best experience." This assessment stemmed from his observation as a technical leader of the team's actual usage patterns, rather than from independent evaluations.

At a broader AI architecture level, DevRev uses GPT-6.1 Sol to power its Computer and Operator products, which are designed for enterprise customer service and internal operations. Bashir described his core criteria when selecting models: correct reasoning capabilities, trustworthy alignment levels, and most critically, latency that meets business requirements. He stated that GPT-6.1 Sol performed well in DevRev's self-built Enterprise-Bench evaluation, which tests models under conditions where data exists in multiple versions, information is scattered across different systems, and access rights depend on user roles. Bashir said the model "strikes the balance perfectly."

It should be noted that this is a customer case study article published by OpenAI, and DevRev's Enterprise-Bench evaluation results have not been independently verified by third parties. Bashir's discussion of developer judgment is his interpretation of a single company's experience and cannot be directly generalized as a universal behavioral pattern for the entire developer community. The decision for Codex to replace Cursor occurred within a North American AI customer service company, whose scale, technology stack, and business scenarios differ from those of other organizations.

Fontes:OpenAI

Pesquisa

ActivTrak Study: Heavy AI Users Work Up to 90 Minutes Longer Daily With More App Switching

Tópico · ActivTrakMaterial
2026-11-04 20:00 GMT+8

Workplace analytics firm ActivTrak has published a new study finding that employees who regularly use AI tools are actually working up to 90 minutes longer every day than those who do not. The research analyzed digital work activity from over 9,000 employees at 467 companies between April 1 and September 25, 2026.

The data shows that the average workday for AI users spanned 8.4 to 9.1 hours, compared to 7.6 hours for non-users. "Heavy" AI users—defined as those using AI tools for more than 45 minutes across at least five sessions a day—also spent an additional 18 minutes working after hours. Furthermore, these users were found to switch between applications more frequently and had shorter periods of "uninterrupted focus" than other workers.

However, the report noted one silver lining: heavy AI users spend significantly less time in meetings, which accounted for just 3% of their day, compared to 5% for non-users. Software engineers and operations managers were identified as among the highest AI users.

"AI may help employees complete tasks faster, but our research shows it contributes to longer workdays, more application switching and shorter focus sessions," said Gabriela Mauch, chief customer officer and head of the ActivTrak Productivity Lab. She advised leaders that the goal cannot simply be AI usage, but rather treating it as a means to measurable gains in productivity and business performance.

Fontes:TechRadar

Pesquisa

Qualcomm Selected by SpaceX to Power Starlink Mobile's Nationwide US Terrestrial 5G Network, Rollout Planned for 2027

Material
2026-11-05 05:32 GMT+8

Qualcomm announced on November 4, 2026, that SpaceX has selected its Dragonwing Cellular Infrastructure platforms to power Starlink Mobile’s nationwide terrestrial 5G network in the United States. The multi-year agreement covers tens of thousands of base station sites, with rollout planned to begin in 2027.

According to the announcement, Starlink Mobile will deploy Massive MIMO radio units using a 64 transmit and 64 receive path (64T64R) configuration. Qualcomm’s QRU100 platform provides highly integrated processing for these radio units, while the QDU100 platform handles high-performance, latency-sensitive Layer 1 processing. This architecture is designed to improve coverage, capacity, and spectral efficiency.

Qualcomm stated that the two platforms are designed to improve performance per watt across both radio and distributed-unit processing. For a nationwide buildout involving tens of thousands of sites, energy efficiency is critical to deployment economics, as power draw compounds quickly at each location and many tower sites face significant power constraints.

This represents a significant win for Qualcomm’s Dragonwing infrastructure portfolio in the US market, which had previously seen more adoption in Asia. If SpaceX can accelerate its terrestrial network deployment, this move positions Starlink Mobile to potentially become a fourth major carrier in the US, competing directly with incumbents like Verizon, AT&T, and T-Mobile.

Fontes:ServeTheHome

Pesquisa

Snap Inc. Reports Surpassing 30 Million Paid Subscribers as AI Features Drive Lens+ Growth

Tópico · Snap Inc.Material
2026-11-05 05:49 GMT+8

US social media company Snap Inc. announced in its third-quarter 2026 earnings release that its total paid subscribers across Snapchat+, Memories Storage, and Lens+ have surpassed 30 million. This represents an increase of 5 million from the 25 million milestone reported in February.

"We surpassed 30 million paid subscribers across Snapchat+, Memories Storage, and Lens+," Snap CEO Evan Spiegel said in a letter to investors. "Innovative AI-powered features have helped Lens+ exceed our expectations, demonstrating how new creative experiences can drive demand for our paid products."

Launched in 2022 at $3.99 per month, Snapchat+ offers exclusive and pre-release features. Snap subsequently introduced additional paid tiers to diversify revenue, including Lens+ ($8.99/month) for exclusive AR experiences and Snapchat+ Platinum ($15.99/month) for an ad-free experience. Since September 2025, users have also been required to pay $1.99/month or subscribe to Snapchat+ to retain Memories beyond a capped free storage limit.

Despite strong subscription growth, Snap faced broader financial headwinds. The company reported 495 million daily active users, falling short of analysts' expectations of 499 million. Shares fell 12% after the release missed ad revenue and user growth estimates. Spiegel noted that while advertising revenue growth remains below target, the company plans to prioritize improving advertiser performance and expanding ways businesses can reach its users.

Fontes:TechCrunch

Pesquisa

OpenAI Global Affairs Head Chris Lehane Pushes AI Policy with New Deal Analogy; Critics Note Proposals Prioritize Promotion Over Constraints

Material
2026-11-05 05:25 GMT+8

With the U.S. midterm elections just concluded and Democrats securing control of Congress, Chris Lehane, OpenAI’s chief global affairs officer, promptly released a lengthy essay attempting to set the tone for the AI policy debate. His core argument posits that AI’s potential to reshape the economy is comparable to the New Deal—the federal programs and reforms enacted in the 1930s after industrialization and the Great Depression shattered the old economic order. Lehane observes that Americans today harbor similar anxieties about losing control over their economic futures and urges politicians to find a compelling third path that enables the public to share in the dividends of AI.

Lehane listed five policy priorities spanning worker protection, education, AI safety, and data center development. On taxation, he suggested incentives for workers to learn AI skills, lowered barriers to capital access for entrepreneurs, and new models for employee ownership and profit-sharing. Regarding data centers, he acknowledged public concerns but emphasized job creation and local tax revenue, warning that politicians supporting a moratorium on data centers could jeopardize America’s technological lead over China, even likening such a stance to the 2002 vote authorizing the Iraq War. On AI safety, he called for enforceable safety standards, independent assessments, public oversight, and a kill switch to restrict or shut down systems if models were to go out of control.

However, Kate Brennan, who leads AI and corporate power policy at the AI Now Institute, pointed out structural flaws in this narrative during an interview with TechCrunch. She argued that Lehane’s rhetoric marks a significant shift in OpenAI’s policy agenda—last year, the company focused on accelerating AI development, reducing regulatory barriers, and expanding infrastructure. Yet, Brennan noted that the essence of the New Deal was establishing worker protections and constraining corporate power, whereas it remains unclear how Lehane’s vision would hold AI companies accountable for the economic disruption their technology might cause.

Brennan’s core criticism is that all these policies presuppose the premise of mass AI adoption and AI benefiting society—a scenario that has not yet materialized. She pointed out that Lehane’s tax suggestions—skill-learning incentives, easier capital access for startups, and employee ownership—are not equivalent to binding requirements ensuring workers share in the wealth generated by AI or requiring companies to cover the costs of displacing workers. A more reasonable approach would involve revising corporate tax rates, providing tax incentives for hiring humans rather than automating jobs, and preventing companies from accumulating excessive market power.

Regarding the stance on safety standards, Brennan labeled it disingenuous. She cited OpenAI’s track record: in 2024, the company opposed California’s AI safety bill SB 1047, which required catastrophic risk assessments and shutdown capabilities; last year, OpenAI also attempted to limit state regulations, arguing that state-level rules could undermine American innovation. Although the company has recently adjusted some positions—supporting several state safety laws and expressing willingness to accept independent evaluators embedded within the company—Brennan believes there is clear tension between these shifts and its historical behavior.

Fontes:TechCrunch

Pesquisa

Forrester Analyst Brian Hopkins: The 'Trust Tax' of AI Budgets and the Three Stages of Return

Material
2026-11-05 04:23 GMT+8

In a Q&A session, Forrester analyst Brian Hopkins proposed an explanatory framework for why enterprise AI budgets continue to swell while returns remain ambiguous. Forrester forecasts that AI will account for approximately 20% of business technology spending in 2027, but Hopkins emphasized that this figure includes infrastructure investments by technology vendors; it does not mean that the average enterprise will devote 20% of its own IT budget to AI.

Unpredictable costs are one of the core issues. Traditional software procurement allows organizations to estimate annual expenses and budget accordingly, but generative AI has introduced billing models based on tokens, usage charges, and credit limits, while agent costs fluctuate with task complexity and required computing power. Hopkins further notes that agentic AI brings an additional hidden expense: companies often expect an agent to directly automate a process, but in reality, they must also test and red-team it, add guardrails and monitoring, maintain a centralized registry, conduct governance reviews, and verify its identity when it interacts with customers or partners. He refers to these extra burdens as the "trust tax" and acknowledges that the industry is still trying to determine how much organizations should expect to spend.

Regarding return measurement, Hopkins distinguishes three time horizons: productivity (short-term), engagement (medium-term), and transformation (long-term). Productivity is the easiest to quantify—for example, whether AI shortens the time employees spend writing customer emails or accelerates ticket resolution—but its impact on the profit and loss statement ranks low. Engagement involves improving how customers and employees interact with the business, such as helping customers obtain answers and complete transactions more easily, thereby increasing transaction frequency and satisfaction. Transformation means redesigning business processes, accelerating product delivery, creating new products, or entering markets ahead of competitors; it carries higher risk but can establish a lasting competitive advantage. Hopkins believes that enterprises focusing only on productivity risk missing the greater opportunities in engagement and transformation, and capturing these benefits requires cross-departmental coordination, shared priorities, and aligned goals.

Hopkins's strategic advice is that enterprises should start with an AI strategy that identifies major business challenges, competitive advantages, and desired outcomes, translate that into measurable targets, and build an investment portfolio overseen by a business executive accountable for results and a portfolio manager responsible for funding. He gives the example of an oil company that might measure performance through cost per barrel, production volume, and safety; AI investments should be tied to improving these metrics. He anticipates that 2027 will be challenging, with some enterprises recognizing the value of AI in engagement and transformation and building business cases, while those that continue to treat AI merely as another productivity tool are more likely to struggle. Many companies may pull back, while a smaller group pushes forward with the right strategy—the full financial impact of this divergence likely won't be visible until 2028 to 2029.

Fontes:AI Business

Pesquisa

Anthropic Says Claude Resolved First Prokaryotic Vault Particle Structure, Cutting Purification Time from 3 Days to 90 Minutes

Material
2026-11-04 08:00 GMT+8

US-based artificial intelligence company Anthropic announced on November 4, 2026, that its biology research lab successfully used the Claude model to resolve the first 3D structure of a vault particle found in prokaryotic cells. Vaults are large, hollow protein shells whose function remains unknown. This discovery offers a simpler system for studying these complex molecules, as bacteria grow faster and are easier to manipulate than mammalian cells.

While vaults are widespread in eukaryotes, their specific role has been an enduring mystery. The Anthropic team noted that while related proteins had previously been identified in prokaryotic genomes, no one had demonstrated that they assemble into complete vault-like particles. Using cryo-electron microscopy, the team determined that this phage-derived vault is built from 96 copies of a single protein called L05, exhibiting structural differences from mammalian versions, such as a shorter, wider shell and a distinct cap-closing mechanism.

Beyond the structural finding, Anthropic highlighted Claude's role in accelerating experimental workflows. According to the company, with Claude's assistance, the traditional vault purification step was reduced from the usual three days to approximately 90 minutes, and the first structural model was ready within 34 hours, compared to the weeks typically required. Claude participated in searching protein databases, designing DNA constructs, optimizing isolation methods, and processing microscopy data.

Despite this progress, the exact biological function of vault particles remains undetermined. Anthropic stated that the study aims to demonstrate how AI can be integrated into every step of fundamental scientific research to speed up discovery. The results have been submitted to the bioRxiv preprint server, with a link to be updated upon publication.

Fontes:anthropic.com

Pesquisa

Richard Windsor: Slowing Growth at Anthropic Casts Doubt on $2 Trillion IPO Valuation

Tópico · Anthropic上市Material
2026-11-04 14:37 GMT+8

In his latest analysis, Richard Windsor, founder of the independent technology research firm Radio Free Mobile, provided a detailed interpretation of Anthropic’s updated S-1 filing, putting forward a core judgment: the $2 trillion IPO valuation is becoming overly optimistic. Windsor previously spent 11 years at Nomura Securities, where he was responsible for equity research covering the global technology sector, before founding Radio Free Mobile to focus on independent research into digital ecosystems.

Windsor highlights a key fact: Anthropic’s Q3 2026 revenue was $16.96 billion, representing a 46.6% quarter-over-quarter increase, but this growth rate is far below the 144.4% seen in Q2. More notably, Q3 operating profit shifted from a $974 million profit in Q2 to a $567 million loss. Windsor attributes this sharp deceleration to a specific mechanism—the retreat of 'tokenmaxxing.' He believes that by the end of Q2, corporate financial controllers realized that token consumption costs were exorbitant and implemented cost controls to prevent massive bills.

The reader value of this explanation lies in linking an seemingly abstract slowdown in growth to observable changes in corporate behavior. Windsor argues that while this state of affairs is perfectly normal, it means that predictions circulating earlier this year claiming Anthropic would exceed $400 billion in revenue by 2027 were nothing more than fantasy. Citing Reuters reports, he notes that Anthropic itself has adjusted its revenue target for 2028 to $190–200 billion, implying that 2026 revenue might be in the $60 billion range and 2027 around $120 billion.

On the cost side, Windsor notes that Anthropic’s total spending commitments have reached $895 billion, with $448 billion due by the end of 2030 and $70 billion in the next 12 months. He considers this figure reasonable—if 2027 revenue reaches $120 billion, compute costs would be approximately $90 billion, leaving $30 billion for other expenses, which could allow the company to break even in 2027. However, Q3 compute costs rose as a percentage of revenue from 48.1% in Q2 to 58.2%. Windsor suspects some price discounts may have been accounted for in this number, which also marks the first quarter where the higher compute prices Anthropic pays to xAI ($37.7 billion per gigawatt) hit the profit and loss statement.

Windsor further points out an accounting treatment issue he suspects: Anthropic offered $600 million in price discounts in Q3 and shifted to cheaper model variants, yet gross margins remained at 85.7%. He speculates that some discounts may have been classified under US GAAP as sales and marketing expenses rather than cost of goods sold, thereby inflating gross margins. He explicitly states that this is his suspicion rather than a confirmed fact.

The competitive landscape is the core basis for Windsor’s valuation judgment. He argues that there is no moat between models; league tables show limited capability gaps among vendors, meaning a price war with OpenAI in 2027 is likely. RFM Research estimates that OpenAI holds a more favorable position regarding compute costs than Anthropic, allowing it to profit at price levels where Anthropic operates at a loss. Windsor wrote: 'For me, this is the key risk to the Anthropic proposition as a public company, and as such I would want a large discount to the valuation to compensate me for bearing this risk.'

Fontes:radiofreemobile.com

Pesquisa

After AI Dismantles the Investment Banking Analyst Pyramid, Winners Redirect Freed Capacity to Deal Sourcing

Tópico · AI投行裁员与转型Material
2026-11-05 00:02 GMT+8

The economic model of traditional investment banks rests on a core number: one. A junior banker supports roughly one deal per year. This is not inefficiency; it is the product of a personnel pyramid built over decades—layers of analysts performing production work to drive deals, topped by senior bankers who own client relationships and judgment. As AI agents become capable junior colleagues, this pyramid is being dismantled.

David Solomon, CEO of Goldman Sachs, described a specific scenario: in the past, when a company prepared for an IPO, a team of about six people would spend two weeks drafting an S-1 regulatory prospectus. Now, AI can complete 95% of that task in minutes. If 95% has become a commodity, the remaining 5% constitutes Goldman’s marginal advantage over competitors accessing the same AI models. This perspective shifts the question from 'how many analysts are needed' to 'what should teams at each level do with the time that is now freed up.'

Deloitte estimates that generative AI could increase front-office productivity at the world’s largest investment banks by 27% to 35% by 2026, adding approximately $3.5 million in revenue per front-office employee. McKinsey found that deal process costs decreased by about 20%, with timelines shortened by 10% to 30%. In February 2026, Evercore’s M&A research desk launched an internal AI tool to automate the creation of precedent transaction decks, allowing VP-level associates to run their own deal teams without analyst support on research tasks. The leverage figures are striking: traditionally, a boutique advisory firm supports one deal per junior banker per year; after AI takes over the analytical layer, the same banker can simultaneously support two to three deals.

However, the author argues that treating this solely as a cost-reduction story—fewer analysts, leaner teams, higher revenue per head—misses both a larger opportunity and significant risks. When you relieve constraints downstream, friction moves upstream. Sourcing is the most critical engine once more deal capacity is unlocked. While 80% of corporate and private equity organizations now use generative AI for target identification and 79% for screening, the average firm still sees only about 18% of relevant deals. Furthermore, 57% of dealmakers report increased senior-level resistance to AI over the past year due to accuracy concerns and fiduciary risk. The bottleneck has moved, but organizations have not yet caught up.

The article must address a tricky piece of evidence: a 2026 benchmark involving 502 current and former investment bankers (from firms including Goldman Sachs, JPMorgan, and Evercore) tested nine frontier AI models against 100 junior banker workflow tasks. Even the best-performing model failed nearly half the grading criteria, and no banker rated any output as ready to send to a client. The failure modes were specific and instructive: fabricated data presented as correctly derived, conflicting figures across slides, and hard-coded values replacing formulas. Only 10% of surveyed bankers believed current AI tools met their needs, with 78% citing excessive errors as the primary limitation.

Fontes:WRITER

Pesquisa

Silicon Valley's 'Leading the Future' Super PAC Fails to Swing 2026 US Midterms, Retains Over $20 Million

Material
2026-11-05 04:41 GMT+8

According to WIRED, Leading the Future, the largest pro-AI super PAC network funded by Silicon Valley tech giants, ended the 2026 US midterm elections with just over $20 million remaining in its accounts. Despite spending approximately $26 million during the primaries and continuing expenditures through the general election, the network failed to secure victories in critical competitive districts, exposing deep divisions within the AI industry regarding regulatory positions.

Launched in August 2025 with a committed war chest of around $100 million, Leading the Future received major donations from Andreessen Horowitz, Palantir co-founder Joe Lonsdale, and OpenAI President Greg Brockman along with his wife Anna. The network aimed to replicate the cryptocurrency industry's success in the 2024 elections by backing candidates opposed to strict AI regulations. However, CNBC's analysis of Federal Election Commission (FEC) filings shows that Anthropic's $40 million commitment to Public First Action, which supports AI safety measures, created direct confrontations between accelerationist and safeguard-focused camps.

In specific races, candidates backed by Leading the Future suffered defeats in several key seats. In South Carolina, Darline Graham, supported by the network, lost with 46.5% of the vote. In North Carolina's 1st Congressional District, Republican Laurie Buckhout, backed by Leading the Future, lost her primary to Don Davis, who was supported by the safety camp. Although the network successfully defeated AI safety advocate Alex Bores in New York's 12th District primary, this tactical victory did not translate into broader strategic advantage. Instead, voter distrust toward Big Tech and data center expansion grew, with NBC exit polls indicating that 56% of voters said AI and data centers played a 'major' role in their voting decisions.

The election results demonstrate that attempts to build a unified AI political machine through massive spending have failed. With Meta returning portions of its political donations and internal employee resistance at OpenAI against Brockman's political activities, Silicon Valley's influence in Washington is undergoing restructuring. Future AI regulatory policy will no longer be dominated by a single industry lobbying force but will depend on the ongoing tug-of-war between accelerationist and safety factions within Congress.

Fontes:WIRED

Pesquisa

AWS Executive in Residence Matthias Patzak: AI Agent Governance Should Be Built at the Platform Layer, Not Fed to Models as Context

Tópico · AWS代理治理平台层Material
2026-11-05 03:59 GMT+8

In a recent interview with AI Business, AWS Executive in Residence Matthias Patzak offered a specific architectural judgment on enterprise AI agent governance: feeding governance rules to models as document context is not best practice; the truly effective approach is to build governance at the platform layer as infrastructure for agent deployment. This distinction stems from his practical consulting experience serving clients in the automotive, banking, and other industries.

Patzak observed that organizations with vague governance rules also have poor AI agent governance. He noted a counterintuitive phenomenon: the financial services industry is actually the most mature in scaling AI, precisely because it has extensive written-down governance rules. These rules can be provided directly to agents as context—but Patzak explicitly stated this is not best practice. While agents do need to know which rules they must follow, the correct approach is to treat governance as infrastructure, a container for agent deployment, setting boundaries at the platform level rather than relying on document snippets within context windows. He considers manufacturing, automotive, and financial services companies to be the most mature in agent governance today.

At the level of multi-agent collaboration, Patzak believes the current best practice is to deploy an orchestration agent to supervise the overall process. He explained that one cannot assume a single agent will never make mistakes; therefore, second and third agents are needed to take on different roles. Role-specific agents require deep expertise but have limited context windows, necessitating an orchestration agent to coordinate the big picture. However, he emphasized that this architecture is currently built by organizations and developers as independent software components and has not yet formed a mature product.

Fontes:AI Business

Pesquisa

US AI Infrastructure Firm Crusoe Debuts 3.5 MW Liquid-Cooled Modular Data Center Unit to Bypass Grid Delays

Material
2026-11-05 03:47 GMT+8

US-based AI infrastructure company Crusoe debuted its new Spark Liquid modular data center unit on November 4, 2026. The unit is designed to support up to 3.5 megawatts (MW) of compute capacity per module and is built specifically for Nvidia GB300 NVL72 and Vera Rubin NVL72 racks. Crusoe stated that this platform will join its existing production line in Brighton, Colorado, with manufacturing scheduled to begin in 2027.

Unlike traditional on-site construction, Spark Liquid utilizes a closed-loop direct liquid cooling (DLC) system to manage heat from high-density GPU clusters. Bryan Schramm, Head of Spark at Crusoe, noted that as rack power continues to rise, air cooling can no longer remove sufficient heat, making DLC necessary. The system recirculates water in a closed loop, minimizing ongoing water usage and eliminating evaporative loss. Schramm emphasized that this factory-built approach addresses three key constraints facing AI infrastructure: power availability, predictability, and "time to token."

Regarding power, large grid interconnections often take years to secure, whereas modular units can leverage available distributed power sources such as on-site generation, small modular reactors, or solar paired with batteries. This allows for faster connection to demand centers. In terms of predictability, factory environments offer more reliable delivery dates compared to construction sites, which is critical for organizations reserving capacity in advance. Additionally, modular infrastructure significantly shortens the time from build to revenue, with Schramm stating that capacity can come online in as little as a few months from a ready site.

Crusoe also announced that its Spark platform will support Nvidia's recently released DSX POD standard, a blueprint for modular infrastructure designed for the Vera Rubin platform. Nvidia designs and validates the reference architecture, while partners manufacture and integrate the units. Crusoe is one of eight container solution providers listed by Nvidia. While specific performance metrics are vendor-reported, this move reflects the industry's shift toward vertical integration and standardized modules to overcome physical and regulatory bottlenecks.

Fontes:DataCenterKnowledge

Pesquisa
Próxima →
Entrar / Cadastrar
LeituraPesquisaRadarFramework de investimento

Lehane also advocated that people should maintain control over major decisions made using AI, rather than handing judgment entirely over to machines. But TechCrunch’s article raised a follow-up question: what does human control actually mean in a society where humans are increasingly required to use AI in education, work, healthcare, and government?

Throughout his essay, Lehane frames AI as a technology that could restore Americans’ sense of agency and economic opportunity. However, he ties the middle-class success of average Americans to participation in an economy increasingly shaped by companies like OpenAI. It remains unclear whether his policies will meaningfully redistribute power or subtly entrench it.

The takeaway for readers is this: when an AI company begins packaging its policy proposals with historical metaphors like the New Deal, the question worth asking is not whether the rhetoric is moving, but whether it contains the core elements of the historical precedent it invokes—the use of state capacity to truly benefit the public and actual constraints on private power. Lehane’s proposals offer concrete pathways for expanding AI adoption and infrastructure but leave glaring gaps in constraining corporate power. This gap itself is key to understanding the trajectory of the AI governance debate.

It is important to note that Hopkins's framework is an analytical model proposed by an analyst, not an independently verified empirical conclusion. The 20% forecast includes vendor infrastructure spending, the specific scale of the "trust tax" remains undetermined, and the division into three stages reflects Forrester's methodology. These boundaries do not affect the explanatory power of the framework, but readers should be aware of its assumptions when applying it to their own budget decisions.

Based on this analysis, Windsor suspects the IPO pricing may be around $1.5 trillion, rather than the $2 trillion discussed in the market. He acknowledges that Anthropic and its backers may be unwilling to accept such a discount, but he chooses to wait and see—'I am still sitting this one out.' It should be noted that Windsor’s assessment of OpenAI’s cost advantage comes from RFM’s own research rather than public data, and his expectation of a $1.5 trillion pricing is personal speculation rather than confirmed market information.

The author’s response is that this is not a verdict on AI in banking, but a verdict on ungoverned AI. Fabrication, inconsistency, and unreviewable outputs are precisely the failures that governed, data-anchored AI aims to prevent. When AI is anchored to a firm’s proprietary data, compliance guardrails govern every generated word, and outputs reflect a banker’s personal style rather than a generic template, the reliability gap narrows. This argument is logically sound, but its evidentiary boundaries must be preserved: the author’s employer, Writer, is the vendor providing such governed AI and compliance guardrails. The claim that 'governed AI bridges the reliability gap' is its commercial proposition and has not been independently verified by the benchmark tests.

The author’s core thesis is that companies winning the next cycle will not use AI leverage to shrink teams, but will redeploy freed capacity toward things AI cannot do: sourcing. The author notes that before joining Writer, they managed multi-asset portfolios and worked with private market distribution teams. From their observation, what distinguishes bankers who source well from those who do not is never the list of prospects, but the conversations. AI can surface targets, score universes, and draft first-pass outreach emails, but it cannot detect management team conviction, read tension in a boardroom, or sense whether a founder is ready to sell or just testing the waters. The advantage lies in building the trust that makes a CEO pick up the phone when your number appears.

After Evercore removed the analyst research layer, internal commentary was sharp: the path from associate to VP became steeper because the mid-level support layer that VPs used to develop on was removed. The author’s answer is not to eliminate those analysts, but to redeploy them into sourcing—the work that builds judgment, relationships, and conviction, which defines a senior banker. An analyst spending their week on deal sourcing research and personalized outreach is developing the exact skills required for the associate-to-VP path; the analyst who is laid off never develops these skills.

The article proposes a 'franchise model' as a mechanism: the best practices of top sourcers are encoded into governed playbooks and scaled across the team, so every associate’s outreach reflects Managing Director standards rather than generic templates. The author states that Writer’s voice feature learns individual bankers’ writing styles—phrasing, habits, and idioms that make an email recognizably theirs. One financial services client noted: if a client can tell they are reading something written by AI, it is offensive. In investment banking, where the product is credibility and senior attention, that standard is even higher. Without built-in compliance guardrails, personalized drafts sit in review queues until bankers give up and send pre-approved templates.

Key boundaries to retain: the persuasiveness of this argument depends on the premise that 'governed AI can indeed bridge the reliability gap,' a premise currently supported only by the author’s employer’s product narrative, lacking independent verification. The benchmark data itself indicates that current AI still fails broadly in investment banking workflows. There is an unverified leap between 'AI can automate the analytical layer' and 'freed capacity can be successfully redeployed to sourcing.' The article’s value lies in posing the correct strategic question—that the bottleneck has moved upstream and layoffs are short-sighted—but betting the answer on the reliability of governed AI is both its core thesis and its most vulnerable point.