2025 September - What I Have Read

Posted on

Substack

How to Handle Visionary Leaders Without Losing the Team - Amy Mitchell, Product Management IRL [Link]

visionary_v_execution

Microsoft announced AI credits for Copilot in Microsoft 365 in January. Salesforce added a new flexible, credit-based model for their AI agent in May. Cursor shifted to credit-based pricing in June (and faced some real pushback from users). Not to be outdone, OpenAI recently replaced seat licenses with a pooled credit model for its Enterprise plans.

― Why everyone’s switching to AI credits - Kyle Poyar, Growth Unhinged [Link]

Companies are transitioning to credit-based pricing models, particularly for AI services, for several key reasons related to managing costs, maximizing profitability, accommodating evolving AI technology, and establishing market standards.

  • The shift to credit-based models is largely driven by challenges related to AI operational expenses and usage patterns.
  • Companies are using credits as a mechanism to transition from flat-rate pricing toward models based on the value delivered.
  • The move by major technology companies validates and standardizes the credit model for AI consumption.
  • Credit models offer flexibility for both vendors and users.
  1. Focus on what you can do. End on an affirmative.
  2. Cite trade-offs.
  3. Get more info to make an informed decision.
  4. Add “because” to share your rationale.
  5. Give the benefit of the doubt.

― Why "'no' is a complete sentence" is dangerous advice - Wes Kao's Newsletter [Link]

How to make your writing C.R.I.S.P. - Dan Hock's Essays [Link]

How To Expand Your Influence Skills - Yue Zhao, The Uncommon Executive [Link]

Shaping the opinions of others, or building influence, is about more than just data and logic. It's about understanding and managing emotions.

Handling your own emotions: Notice and reflect on what is driving your actions, such as fear, and then name it. This helps you move forward with clarity and confidence.

Leading others through their emotions: When you want to get buy-in for your ideas, help people process their emotions. You can do this by creating a space that welcomes emotions, validating their concerns, and then shifting their focus to what they can do to move forward.

The Hidden Rulebook of Corporate Politics (and How to Use It to Your Advantage) - Gaurav Jain, The Good Boss [Link]

I have to review this article regularly.

The moment you stop believing in the corporate fiction is the moment you can start using it. Once you see it as infrastructure rather than identity, as a resource rather than a calling, everything shifts.

Your corporate role doesn't need to be meaningful. It needs to be useful. Useful for building skills, for funding your real projects, for buying time while you figure out what matters to you.

The death of the corporate role isn't a crisis. It's freedom from having to pretend your spreadsheet about spreadsheets is your life's work.

― The death of the corporate job - Alex Mccann, Still Wandering [Link]

Good piece.

Articles and Blogs

President Trump, Tech Leaders Unite to Power American AI Dominance - The White House [Link]

The August jobs report has economists alarmed. Here are their 3 top takeaways. - CBS News [Link]

The August jobs report is raising concerns among economists due to several alarming trends. Employers added only 22,000 nonfarm jobs, which is significantly lower than the 80,000 jobs that analysts had forecast. The unemployment rate also rose to 4.3%, the highest level since October 2021.

The three top takeaways are

  1. The job market is stalling
  2. Job growth is at its lowest level in 15 years
  3. The federal reserve will likely cut interest rates

The Recession is Already Happening for Many Americans - Bloomberg [Link]

Read the text messages between Charlie Kirk accused and roommate - BBC [Link]

U.S. Investors, Trump Close In on TikTok Deal With China - Raffaele Huang, Lingling Wei, Alex Leary, The Wall Street Journal [Link]

The near-finalized framework of a deal between the U.S. and China concerning the popular social media application TikTok, involves creating a new U.S. entity to manage the app’s American operations, with an investor consortium, including Oracle, taking a roughly 80% controlling stake, which satisfies a recent U.S. law regarding foreign ownership. A key component of the agreement is the establishment of American control over user data and the crucial content-recommendation algorithms, although they will be based on technology licensed from TikTok's Chinese parent company, ByteDance. Furthermore, the article notes that President Trump has delayed the TikTok ban until December as negotiations conclude, signaling the resolution of a multi-year national security dispute over the app's influence in the U.S. Both Chinese and American officials have reached a basic consensus on the terms, which also include Oracle managing U.S. user data at its facilities in Texas.

Google brings Gemini in Chrome to US users, unveils agentic browsing capabilities, and more - TechCrunch [Link]

NotebookLM - Jason Spielman [Link]

Tesla Dojo: The rise and fall of Elon Musk’s AI supercomputer - TechCrunch [Link]

Dojo was a custom-built supercomputer intended to be the cornerstone of Tesla's AI ambitions, specifically for training the neural networks of its Full Self-Driving (FSD) technology and humanoid robots.

The primary strategic reasons cited for the project's termination include:

  1. the strategic pivot to AI6 chips. The AI6 chip is Tesla’s new strategic bet on a chip design intended to scale across FSD, Tesla’s Optimus humanoid robots, and high-performance AI training in data centers
  2. moving away from hardware self-reliance. Dojo was intended to reduce reliance on expensive eand difficulty-to-secure Nvidia GPUs, but Tesla is now "going all-in on partnerships" with major chip providers, including Nvidia, AMD, and Samsung (which will build the AI6 chip)
  3. technological and compatibility hurdles. Dojo’s design, based on proprietary D1 chips, faced inherent technological challenges related to integration with the broader AI ecosystem
  4. internal competition and redundency. In August 2024, Tesla began promoting Cortex, described as the company’s "giant new AI training supercluster" being built at Tesla HQ in Austin. Cortex was later deployed at Gigafactory Texas.

You are by default a product leader, navigating product directions with data.

Data scientists at Meta don’t just analyze data — they transform business questions into data-driven product visions that help building better human connections.

The most successful data scientists that I’ve worked with not only excel at adapting their approach to the specific data-problem quadrant they’re operating in, but also are effective in working with Cross-Functional partners to drive collaboration pushing product strategy development forward.

With Product Managers:

  • Speak in terms of business problems, not data techniques
  • Help PMs translate intuition into testable hypotheses
  • Co-create metrics frameworks that balance short and long-term objectives

With Engineering:

  • Bridge implementation and insight by understanding technical constraints
  • Design analytics requirements that respect engineering resources
  • Create feedback loops that allow for continuous improvement

With Design/User Researchers:

  • Humanize data insights through collaborative storytelling
  • Provide quantitative context for qualitative user research
  • Partner on creating experiences that naturally generate valuable data

Deb Liu, former VP of Meta, highlighted in herproduct strategy framework: “a great product strategy is opinionated, objective, operable, and obvious.” Data scientists are uniquely positioned to help product teams achieve these qualities through:

  • Opinionated: Grounding strategic choices in data-backed insights
  • Objective: Bringing analytical rigor to opportunity sizing and risk assessment
  • Operable: Creating measurement frameworks that make execution tractable
  • Obvious: Revealing patterns that make the path forward clear to all stakeholders

― Meta’s Data Scientist’s Framework for Navigating Product Strategy as Data Leaders - Medium [Link]

Quadrant 1: The Pioneer (Low Data, Broad Problem)

  • Strategic Approach:

    1. Identify North Star metrics
    2. Design metric framework
    3. Conduct minimum viable analytics, quasi-experiment
    4. Generate data insights to identify problems and guide early decisions
    5. Create product strategy to drive measurable improvements

    Note: a good strategy decides which problems to prioritize in solving as well as those we choose not to solve.

    Best Practice: narrowing the problem space through structured discovery.

  • Collaboration among design, PM, and XFN

    1. Define (north star) metrics
    2. Translate business questions into testable hypotheses
    3. Use analytics to yield insights

Quadrant 2: The Craftsperson (Low Data, Concrete Problem)

  • Strategic Approach:

    1. Design targeted data collection aligned to the specific problem
    2. Develop creative measurement frameworks that work with sparse data
    3. Leverage analogous data from similar contexts

    Note: focus on setting clear learning milestones rather than promising specific outcomes. The goal is to systematically reduce uncertainty around a concrete problem with iterative data learnings to update our beliefs.

Quadrant 3: The Explorer (High Data, Broad Problem)

  • Strategic Approach:

    1. Pattern recognition at scale to identify unrecognized opportunities (e.g., opportunity sizing model, gap analysis framework)
    2. Segmentation and clustering to create structure in an ambiguous space (e.g., segmentation model)
    3. Insight translation that transforms data patterns into business narratives

    Note: structure the problem space through data, allowing the product team to move from broad exploration to targeted opportunities. The role is to transform overwhelming data into clear strategic choices for your product partners.

Quadrant 4: The Optimizer (High Data, Concrete Problem)

  • Strategic Approach:
    1. Metric deep-dive and monitoring
    2. Analytics modeling that uncovers non-obvious optimization opportunities (e.g., multi-armed bandit system, analytics framework, feedback loop)
    3. Continuous learning systems that adapt as conditions change

Best Practices for Developing a Product Strategy - Deb Liu [Link]

A New Ranking Framework for Better Notification Quality on Instagram - Engineering at Meta [Link]

While existing machine learning (ML) models optimize for high engagement, they can result in repetitive and potentially "spammy" notifications, leading users to disable them. To combat this, the new framework applies a multiplicative penalty to notification scores based on their similarity to recently sent ones, using criteria such as author and product type. This strategy has successfully reduced notification volume while increasing engagement rates by ensuring a more varied and personalized mix of content.

The methodology begins with the existing machine learning (ML) models, which calculate a base score for notification candidates based on factors like the probability of a user clicking (Click-Through-Rate or CTR) and time spent. The new framework introduces a diversity layer on top of these existing engagement ML models.

The methodology involves the following steps:

  1. Evaluation of Similarity: The diversity layer evaluates each notification candidate's similarity to recently sent notifications across multiple dimensions, such as content, author, notification type, and product surface.
  2. Application of Penalties: The system applies carefully calibrated penalties, expressed as multiplicative demotion factors, to downrank candidates that are too similar or repetitive to recent notifications.
  3. Re-ranking: The adjusted scores (base relevance score multiplied by the demotion factor) are used to re-rank the candidates.
  4. Selection: The final selection process uses a quality bar to choose the top-ranked candidate that successfully passes both the ranking and diversity criteria.

Within the diversity layer, the methodology is mathematically implemented using a multiplicative demotion factor applied to the base relevance score:

\[\text{Final Score } (S(c)) = \text{Base Ranking Score } (R(c)) \cdot \text{Diversity Demotion Multiplier } (D(c))\]

Key aspects of this calculation include:

  • Demotion Multiplier (\(D(c)\)): This is a penalty factor where the value falls within the range of 0 to 1 (\(D(c) \in\)), reducing the score based on similarity to recently sent notifications.
  • Similarity Signal: To calculate \(D(c)\), a similarity signal (\(p_i(c)\)) is computed for a set of semantic dimensions (e.g., author, product type) using a maximal marginal relevance (MMR) approach.
  • Binary Baseline: In the baseline implementation, the similarity signal \(p_i(c)\) is binary: it equals 1 if the similarity exceeds a predefined threshold (\(\tau_i\)), and 0 otherwise.
  • Flexible Control: The methodology defines the final demotion multiplier using adjustable weights (\(w_i\)), which control the strength of demotion for each respective dimension.

The State of AI in Financial Services in 2025 — views from our front row seats - Peter Hung, Illuminate Financial [Link]

illuminate_financial_ai_apps

Agentic Design Pattern [Link] [code]

Product Roadmap Examples - Janna Bastow, ProdPad [Link]

The best roadmaps aren't checklists; they tell a story about why something is being built. They show how short-term initiatives connect to long-term strategic goals.

The Now, Next, Later Framework is a core pattern, reflecting the reality of uncertainty. Now initiatives are tight, concrete, and focused on current goals (e.g., MVP launch). Next initiatives are more exploratory bets. Later initiatives are deliberately fuzzy, long-term aspirations that signal intent without making firm promises.

Effective roadmaps frame initiatives as problems to solve and tie them to clear outcomes and business objectives (e.g., "reduce onboarding friction" instead of "ship a new login flow"). This keeps the team flexible and focused on results.

There is no one-size-fits-all roadmap. A startup's roadmap is about survival and proving a hypothesis. A scale-up's roadmap is about smoothing friction and deepening engagement. A hardware roadmap must account for manufacturing cycles, while a mission-critical one must prioritize compliance and security.

Expanding economic opportunity with AI - OpenAI [Link]

GenAI Doesn’t Just Increase Productivity. It Expands Capabilities - BCG [Link]

A point made around 'reskilling': While GenAI can immediately boost a worker's aptitude for new tasks, it does not necessarily "reskill" them in a traditional sense. The study found that participants were able to perform complex data-science tasks with the help of GenAI, but they did not retain the knowledge or skills gained after the tools were taken away. The article refers to GenAI as an "exoskeleton" that enables workers to do more, but does not intrinsically change what they have learned.

Building Etsy Buyer Profiles with LLMs - Isobel Scot, Etsy Code as Craft [Link]

Non-Obvious Tips for Landing the Job You Want - Deb Liu [Link]

Seven Non-Obvious Strategies

  1. Never rely only on online submission: Avoid the "digital dustbin" by finding an alternative path in, such as a referral, connection, or direct reach-out.
  2. Ask for advice, not a job: Sincerely seek guidance on entering a field or company, as people are often generous and may uncover new opportunities for you.
  3. Give them a reason to say yes: Counter the process of finding reasons to say no (misspellings, poor grammar) by providing human connection points like shared alma maters, hobbies, or passions to hack affinity bias.
  4. Find the “you-shaped hole”: Seek roles where your unique skills, experience, or passion make you the best bet, demonstrating you can "hit the ground running on day one".
  5. See the world through the hiring manager’s eyes: Hiring managers prioritize managing risk because a bad hire is costly. Your job is to close the asymmetry of information, prove you are a "sure bet," and show you are a great return on investment.
  6. Do the job before you get the job: Demonstrate initiative by acting like an employee; use the product, talk to customers, and bring specific ideas or prototypes to show you want this job.
  7. Tailor your resume (and your story) for the role: Treat your resume as a "living document" to tell a specific story, reframing factual experiences to align with the target role and "speak the language of the hiring company".

American Express is Accepted at 160 Million Merchants Around the World; Since 2017, Amex-Accepting Locations Have Increased by Nearly 5x - Business Wire [Link]

Hallucinations persist partly because current evaluation methods set the wrong incentives. While evaluations themselves do not directly cause hallucinations, most evaluations measure model performance in a way that encourages guessing rather than honesty about uncertainty.

― Why language models hallucinate - OpenAI [Link]

Several clarifications mentioned in this blog:

  1. Accuracy will never reach 100%
  2. Hallucinations are not inevitable. Language models can choose to abstain when uncertain. Abstaining (indicating uncertainty) is better than providing confident, incorrect information, aligning with the core value of humility
  3. Avoiding hallucination can be easier for a small model to know its limits. Being "calibrated" (knowing its confidence) requires much less computation than being accurate
  4. To measure hallucinations, all of the primary evaluation metrics need to be reworked to reward expressions of uncertainty. Hallucination evals have little effect against hundreds of traditional accuracy-based evals that punish humility

How to Think About GPUs - How to Scale Your Model [Link]

'A Systems View of LLMs on TPUs'

Anthropic Economic Index report: Uneven geographic and enterprise AI adoption - Anthropic [Link] [PDF]

Key findings:

I. Adoption Speed and Shift to Delegation

  • AI adoption is occurring at an unprecedented speed, reaching in two years the adoption rates that took the internet around five years. In the US, 40% of employees report using AI at work, doubling the rate from two years prior in 2023.
  • Usage patterns on Claude.ai show a net shift toward delegation (automation). The share of "Directive" conversations, where users delegate complete tasks, jumped from 27% to 39%, meaning automation usage now exceeds augmentation usage for the first time.
  • There is sustained growth in knowledge-intensive tasks like education and science. In coding, there is a net shift of 7.4 percentage points toward program creation and away from debugging, suggesting models have become more reliable.
  1. Geographic Concentration and Inequality Risk
  • AI usage is highly geographically concentrated and correlates strongly with income across countries. A 1% increase in GDP per working-age capita is associated with a 0.7% increase in Claude usage per capita.
  • Small, technologically advanced economies lead in per-capita adoption, with Israel (7x expected usage) and Singapore (4.57x expected usage) being top examples.
  • Low-adoption countries are more likely to delegate complete tasks (automation), while high-adoption countries tend toward greater learning and collaborative iteration (augmentation), even when controlling for task mix.
  • Current usage patterns suggest that AI benefits may concentrate in already-rich regions, potentially increasing global economic inequality.
  1. Enterprise Automation and Deployment Bottlenecks
  • Enterprise usage via the 1P API is predominantly automation-dominant, with 77% of business uses involving automation patterns (delegating tasks programmatically), compared to about 50% for Claude.ai users.
  • Business deployment is largely price-insensitive. Model capabilities and the economic value of automation appear to matter more than cost, as higher-cost tasks tend to have higher usage rates.
  • For complex tasks, deployment is constrained by the access to information rather than just model capabilities. Companies face a bottleneck requiring costly data modernization and organizational investments to centralize the contextual information needed for sophisticated AI use.

Papers and Reports

NCRI Assassination Culture Brief - NCRI and Rutgers University [Link]

Political violence targeting figures like Donald Trump and Elon Musk is becoming normalized. The report's key findings are based on a survey and social media analysis. Main points:

  1. Growing justification for violence
  2. The rise of "Assassination Culture"
  3. Social Media as an Amplifier

YouTube and Podcast

Trump Takes On the Fed, US-Intel Deal, Why Bankruptcies Are Up, OpenAI's Longevity Breakthrough - All-In Podcast [Link]

Elon Musk on DOGE, Optimus, Starlink Smartphones, Evolving with AI, Why the West is Imploding (All-In Summit) - All-In Potcast [Link]

Inside the White House Tech Dinner, Weak Jobs Report, Tariffs Court Challenge, Google Wins Antitrust - All-In Podcast [Link]

预测未来，还是操纵未来？Polymarket的崛起之路与争议【深度】- 硅谷101 [Link]

E206｜临近机器人GPT-3时刻，具身智能开源模型的加速演进 - 硅谷101播客 [Link]

Charlie Kirk Murder, Assassination Culture in America, Jimmy Kimmel Suspended, Ellison Media Empire - All-In Podcast [Link]

How AI Will End 50 Years of Economic Stagnation, with Tyler Cowen - Working Intelligence [Link]

H-1B Shakeup, Kimmel Apology, Autism Causes, California Hate Speech Law - All-In Podcast [Link]

Social Media

If you believe free speech is for you but not your political opponents, you're illiberal.

If no contrary evidence could change your beliefs, you're a fundamentalist.

If you believe the state should punish those with contrary views, you're a totalitarian.

If you believe political opponents should be punished with violence or death, you're a terrorist.

― J.K. Rowling [Link]

To build an AI native product, a PM needs mastery of the following - vision, opinionated UX design - model intuition to extract max value - ability to go from pixels -> evals -> hill climb - understanding of agentic flows - tools, context, safety guardrails - deep user understanding - lot more than previously because of the nature of agents

― AI PM mastery is a rare skill - Madhu Guru [Link]

The Systems Thinker's Blindspot - Shreyas Doshi [link]

system_thinker_blindspot