Rajeeva, co-founder and CTO of Acceldata, argues that agents can take over much of enterprise data management, from cataloging and quality checks to repairing broken pipelines. That works only if the AI is given context about data it was never trained on, and it stays bounded: an expert still approves the result. Acceldata, founded in 2019 by four former Hortonworks engineers, has raised over $100 million across three rounds. It runs engineering largely from India, with staff in the Bay Area and Canada and a team forming in London.
Agentic data management. Rajeeva says data catalogs, MDM and curation have improved only incrementally for 15 to 20 years and are mostly forms and clicking through defined workflows, which agents can automate. ADM now has close to 10 agents, such as a data-catalog agent that knows a table's columns, types, refresh cadence and owner. They are distilled from customer conversations because banks, insurers and healthcare firms have different problems. Powering them is the xLake reasoning engine. Just as Google's index lets a chatbot search the web, xLake gives AI tools to query petabytes across Snowflake, AWS, Azure and on-prem clusters, while handling scale and column-level permissions.
Self-healing pipelines. An agent detects an anomaly in logs, clones the Spark code, adds error and metadata context and redeploys the fix to the cluster, replacing the bug-report, fix and CI cycle. Rajeeva says Gemini 3 one-shots about 90% of what you want. He concedes proprietary Informatica, HVR or Oracle stored-procedure pipelines are hard, and predicts pipelines will move from drag-and-drop to code.
Humans at the gate. Autonomy, he says, means doing most of the work unsupervised, perhaps a week compressed into an hour, while experts approve, reject or demand rewrites, as in code review. His MCP demo, which built a GDPR compliance-scoring agent and dashboard, got about 60% of the work done, maybe 80% with prompting. Enterprises are slow to decide but afraid of falling behind, so Acceldata builds for where they will be.
Architecture and sprawl. Data sprawl is legacy on-prem data plus Snowflake or Databricks plus a Google cloud deal, all needing to be found and joined. Unlike metadata-only players, Acceldata deploys a data plane in the customer's environment over mTLS and can validate up to 100 million rows. It builds on Kubernetes and Spark and publishes its ODP data platform as open source. Customers are enterprises such as Nestlé, not startups.
Leading engineers. Rajeeva says a CTO must take the customer's view, because customers need wins to renew a proof of value and product managers focus on features. Engineers, who studies suggest are productive only 3 or 4 hours a day, quit without a mission. He keeps them with engineer-led projects: ODP, an AI agent that fixes vulnerabilities, and an in-house data-center stack replacing VMware and OpenStack. He hires senior people on whether they can build. Coding interviews are now face to face because remote LLM cheating is too hard to detect, though he doubts the approach will scale.
Looking back on their Ignite session, Hanselman and Mark Russinovich, CTO of Azure, conclude that a talk with almost no instructional content can still earn its slot if it is honest about what it is and stacks real technology into something delightful. A lot goes on behind a successful tech demo, especially under a tight time limit, and the small thing that hurt them most was not seeing the next slide.
Soft talks and credibility. Mark recalls his first big non-technical talk, on Azure strategy and the cloud landscape around 2015: stuck on altitude until the night before, he made five slides and talked for 75 minutes. Hanselman says he once turned an invitation to speak on personal branding into a talk calling it nonsense, and argues everyone has a perspective worth a talk. Mark counters that audiences need a reason to listen, and they agree a speaker needs a hook that establishes credibility.
Pitching the Ignite talk. Hanselman adapted his hour-plus Portugal talk, "the great circle from Altair to Azure", into 45 minutes, with Mark supplying heavier Azure demos. Mark doubted it until the night before, asking what anyone would take away, and now concedes he was wrong. Their fix was to tell the audience at the start to lower its expectations.
The demos. Mark paints games onto Task Manager on Azure's largest VM (24 TB of RAM, 16 sockets, a processor count captioned as both 792 and 1,792) by modulating per-processor load. The history runs from flipping bits on the Altair, through Mark's 1985 magazine machine-code program and Microsoft BASIC under CP/M, to early websites and Mark's 60,000-subscriber newsletter. The finale piped the Altair's lights by camera, UDP and websockets onto the big VM at about 5 frames per second, alongside Zork on the Altair and Azure Cloud Shell and a swarm of 100 GPT-4o players.
What went wrong. The conference's machine switcher cost them presenter view, so neither could see the next of about 30 slides. They fell behind midway, then the risky final segment ran fast because everything worked. A live hunt through ZoomIt's timer colour settings turns up an inconsistency Mark concedes.
In their fifth annual predictions episode, St. Lawrence, founder of the Human-Machine Collaboration Institute (HMCI), and the host end on a shared view. Scaling has run short of data and clear direction. The next gains will come from specialized models, continual learning, new research and spatial data. Meanwhile intelligence keeps getting radically cheaper. The host is series editor of her new Pearson book, Becoming an AI Orchestrator, and says he urged her to write it.
HMCI and the book. St. Lawrence's TEDx talk urged asking hard questions about merging with AI and the nature of consciousness. HMCI, an independent research firm working toward a unified theory of consciousness, built an intelligent revenue reinvestment (IR2) model. It channels money from energy and data-center projects into local upskilling, research and infrastructure, and she says Nvidia is partnering on a flagship in California, with more cities planned. The book's premise is moving from musician to conductor of AI agents, which she says let HMCI run as a team of three.
Scoring 2025. Agentic AI dominated so thoroughly that audiences now tire of it. The host adds that agents' task length doubles every 7 months at about 50% reliability, pointing to a full workday by end of 2026. AI reached devices such as AirPods' live translation, though she found CES appliances a poor use of it. AI science was mixed: a materials-discovery paper was retracted, but DeepMind's GNoME project impressed. Enterprise monetization remained the issue, with projects stuck at MVP. The host stresses that the same capability now costs about 100 times less than a year ago. A LinkedIn report bore out her call that AI engineering skills would outgrow data science, which she frames as the data-scientist "unicorn" splintering into a family of roles.
The 2025 awards. For wow moment, she picks Nano Banana's accurate room redesigns. He picks Ed Donner's four-agent crew building a working simulated trading platform over a 90-minute lunch at ODSC East. Google is comeback of the year again, for its integration across Search and G Suite and the data users already entrust to it. For disappointment she names enterprise agents, as in her Substack post "Agents of Disappointment", and he names Apple Intelligence. The host also credits xAI's Grok, which he says was statistically tied with Gemini 3 Pro on LM Arena. Overall winners are Nvidia and "vibes", with GPT-5 disappointing and nobody anticipating a next leap. The host notes coding leads because its results are verifiable and its training data can be simulated.
Predictions for 2026. First, specialized AlphaFold-style domain models, since general models have hit data limits. Second, continual and nested learning, after a new Google paper. Third, a return to research: she cites Ilya Sutskever's recent interview as showing labs lack a clear path. The host argues sample-efficient learning, like a toddler inferring an adult's intent, is the missing breakthrough, but doubts it lands in 2026. Fourth, physical AI and spatial intelligence, valued as new data, with 2027 still "the year of the robot". Fifth, AI ops as a new job, akin to DevOps around 2010.
Vjekoslav Krajačić, the developer behind File Pilot, ends up holding that native apps could be far faster and leaner if more developers built them the way games are built, though Krajačić doubts the next generation will learn to, with AI and deep stacks pulling programmers further from the hardware.
From quiet project to public beta. Krajačić started File Pilot in 2022 after leaving the game company they had run for over a decade, saw a gap because most file managers dated from the Windows 95 era, worked incognito for two or three years, had a few hundred alpha users, and drew wide attention with the public beta in early 2025.
Built like a game. Hanselman asks why games run at 100 frames a second while apps stutter on resize. Everything is custom: OpenGL rendering (a DirectX move is planned over driver bugs), an immediate-mode UI that redraws each frame, and Krajačić's own standard library. Hanselman's analysis found a 2.3 MB executable with 14 DLL dependencies and 312 imported functions.
Where it hurts. The worst problems are in old Windows APIs: the context menu stays single-threaded because shell extensions break off the main thread, and drag and drop still has unsolved bugs. Hanselman offers Microsoft contacts, and live finds network-share handling wanting.
AI and the next generation. Krajačić uses ChatGPT only as a better search engine and then verifies against docs, calling AI-written code dangerous for people who can't spot hallucinations. Hanselman agrees, saying seniors get boosted and learners slowed.
What's next. As recorded: free beta at 0.6.3, a $40 individual perpetual license on pre-order, exit from beta early to mid next year, then Unicode rendering, drag and drop from zip archives, possible macOS and Linux ports, and one more hire.
Loris Cro, VP of community at the Zig Software Foundation, ends up presenting Zig as a modern C rather than a modern C++: a pre-1.0 language whose real advantage is working intimately with existing C and C++ code, run by a small nonprofit that keeps big tech off its board.
Why Zig exists. Cro came to Zig from bioinformatics, needing to teach Redis module writing without C's intimidation, and argues C and C++ carry 40-plus years of quirks, such as OpenCV bindings that compiled only because of include order. Zig, at 0.15.2, promises no backwards compatibility, which Cro treats as freedom to fix mistakes.
Against Rust. Hanselman framed Rust as optimising for memory safety and Zig for control and explicitness; Cro agreed, adding that Zig cross-compiles C and C++ dependencies, which Rust leaves to the user. Where Rust's guarantees stop at safe Rust code, Zig adds type-system detail for C interop and defer cleanup, behind Cro's claim that C libraries are easier to use from Zig than from C.
Comptime. Asked how comptime differs from #ifdef, Cro points to lazy evaluation: unresolved branches are not semantically analysed, so per-OS code needs no guarding, and types can be queried at compile time. The cost is that errors in uncalled code go unnoticed.
Governance and ecosystem. The 501(c)(3) refuses big-tech board seats and avoids reliance on one donor. It left GitHub for Codeberg mainly over GitHub Actions problems, preferring nonprofit services as more stable. Bun began with one person asking questions on Discord and, Cro notes, has since been bought by Anthropic; TigerBeetle allocates memory only at startup.
extern "C" block, and only compiled for Go
because an include guard had already fired earlier in the build.errdefer-style cleanup. Hanselman framed Rust as optimising for
memory safety and Zig for control and explicitness, and Cro agreed.Roese, Dell Technologies' global CTO and chief AI officer, returning to give his annual predictions, argues that AI pays off only when it is disciplined and narrowly targeted. Dell's evidence, as he gives it, is a fiscal year in which revenue grew $10 billion while absolute cost fell 4%. His 2026 predictions follow from that: governance, a new knowledge layer, agents as four-part software systems, resilient AI factories, and sovereign AI put to wider use.
Agentic AI and agent washing. Roese says his 2025 prediction was that "agentic" would be the word of the year. He defines it as the digitization of skills: an autonomous system that reasons and executes while the human supplies intent and validation. He calls the term wildly overused and says many "agents" are tools that bring autonomy to people. The host cites the METR finding that the length of human tasks AI can do at 50% accuracy doubles every seven months, now about 2–4 hours. Roese adds a second dimension: agent ensembles take on composite work that spans people and organizations.
MCP and slop. MCP is a tool and data-access standard, not an interworking protocol. Roese says it is early, lacks security, and is dangerous if used carelessly, so Dell centralizes its MCP servers under its unpublished internal agentic guidelines. The host raises The Economist's word of the year, "slop". Roese answers that Dell runs fewer than about 30 use cases with ROIs of 10:1 to 30:1, and that it is "better to do nothing than to do chaos."
ROI. Roese calls the revenue-cost decoupling a first in Dell's 41 years, a mix of people, process and technology. He repeats his rule that every project needs a hard ROI metric. He cites a report he attributes to McKinsey calling generative AI a "BS work detector."
Governance and data. More than a thousand uncoordinated jurisdictions make external governance "a disaster". His advice is to avoid heavily regulated use cases first. He then describes a knowledge layer of vector databases, knowledge graphs and memory, sitting between systems of record and compute. It is always warm, and agents write to it as well as read from it, so it should live near compute at the edge.
Agents, factories, sovereignty. Dell's agent has four parts: an LLM, a knowledge graph, MCP and A2A. In Dell's studies, agents sharing a knowledge graph and told to value consensus ignored a hallucinating peer. Roese says this mitigates hallucination but does not cure it. About 3,000 customers are building AI factories with Dell, and none are yet resilient. Because agents rarely touch primary data, he says resilience priorities and backup models must change, ransomware protection included. He predicts every country will have a sovereign AI strategy, with uses beyond government such as critical-infrastructure backup and government-certified agents. He says quantum is "just accelerating." The host discloses that he and Roese recorded a television ad together for Dell's AI factory with Nvidia.
Lafeuille is a senior data scientist at Medidata Solutions who posts as "The Fit Data Scientist". Her position is that burnout in data science comes from overwork and a career pointed the wrong way, and that recovery has to be managed with the same care as work.
From burnout to a reset. In finance in New York she worked too much at the wrong thing. She switched to life-sciences data science and hired a fitness and lifestyle coach. She had been working out almost daily, overtrained and underfed, and now trains four times a week around her work and pickleball, keeping the schedule flexible.
Fuel and rest. She eats one gram of protein per pound of body weight and puts most of her carbs around workouts so she doesn't crash mid-work. Rest matters most: eight hours of quality sleep, reading before bed, short phone-free walks to downregulate, and a five-minute brain dump into a notebook before bed.
Her work and training. At Medidata, which makes clinical-trial software for pharmaceutical companies, she began on immuno-oncology analysis and now builds the backend data science. She uses R and Python and favours R for biostatistics. Her Columbia master's in management science and engineering included company projects that built her portfolio. She also credits ChatGPT and similar tools with getting her past perfectionism about writing in English.
Clemm, Dropbox's VP of engineering, who leads Dropbox Dash and earlier led about 400 engineers at Uber, argues that the bottleneck in workplace AI is context, not model capability. The fix, in his view, is to ingest and index all of a team's work content ahead of time and give the model exactly what it needs and nothing more. He makes the same case for organizations: leaders must stay technical and work on the inputs of team productivity.
Universal search and multiplayer AI. Clemm describes Dash as the "2025 version of Dropbox": search, chat and then agentic features across files scattered over Slack, Google Docs, Jira and dozens of browser tabs. He says traditional enterprise search is broken because workers face 12 to 20 separate search bars. His "multiplayer AI" is not multiple agents but AI that knows a whole team. It rests on a graph of normalized content and people, and on collaborative products such as Stacks. He flags prompt injection and data exfiltration as the security problem with AI browsers from Perplexity and OpenAI, and argues that pre-ingesting content avoids crawling at query time.
Work slop. He attributes slop to pressured, goal-less rollouts. AI output should be treated as a first draft. His other remedies are hybrid lexical and vector indexes (BM25 being "still the workhorse"), internal evals, and training led by high performers. He adds that coding agents speed only one stage of software development, so bottlenecks move upstream to requirements and design.
Context engineering. Clemm says effective context windows fill at 100K to 200K tokens despite 1–2 million advertised. Dash's accuracy fell as tools were added, so it moved to a single super tool. He says Cloudflare and Anthropic have published similar approaches. Good context engineers, he argues, come from constrained-systems work (embedded software, UAVs, lossy networks), from search and recommender systems, and from evals.
RAG and side projects. RAG has moved from single-shot vector retrieval to hybrid and agentic retrieval, as coding agents exploring a codebase show. His roughly 100-line open-source answer engine grew out of a fantasy-football preview generator built on GPT-3 around 2021–2022 and was inspired by Perplexity. It continues as Yattle.ai, which routes queries between models. He argues that citations let an answers engine "prove its work."
Leadership. Past Dunbar's number (about 150), culture has to be replicated through leaders, small-group AMAs and a weekly check-in he began at Uber in the pandemic and still runs. He prefers fixing inputs (goals, strategy, org structure, seniority mix, operational rigor, tooling) to counting PRs. His "think like a farmer" point is that teams need resilience against whiplash from constantly leapfrogging frontier models. Over-reliance on A/B tests, he says, slows big companies down and erodes conviction.
Reviewing 2025 on the Lightcone podcast, the YC partners conclude the AI economy has stabilized into distinct, durable layers — models, applications, and infrastructure — each with a workable playbook, even as the details underneath shifted more than anyone expected.
A changing of the guard among models. Diana reports that in YC's Winter 2026 application data, Anthropic overtook OpenAI as founders' top API choice, after OpenAI held over 90% share in early batches and Anthropic hovered near 20-25% through most of 2024-25 before a recent surge past half. She credits Anthropic's internal focus on coding performance, plus a possible "bleed-through" from founders' own coding habits. Gemini has also risen sharply, from single digits to about 23%.
Model choice as arbitrage, not loyalty. Partners describe both consumers and founders treating models as interchangeable, hand-picking whichever performs best per task — one series-B company reportedly routes Gemini 3 for context engineering into OpenAI for execution — likening it to swapping chip architectures between Intel and AMD.
Reframing the bubble question. Drawing on economist Carlota Perez's installation-versus-deployment framework, the partners argue the massive current capex — data centers, power constraints prompting ideas like space-based data centers and fusion startups — sits on big incumbents' balance sheets, not founders', so a pullback wouldn't strand startups; the resulting compute glut instead sets up a deployment phase for application-layer companies, comparable to excess telecom fiber later enabling YouTube.
Fine-tuning's fading moat, and hiring. They note early fine-tuned "AI-headed" companies got overtaken once frontier models leapt ahead, though newer domain-specific fine-tunes on open-source bases can still beat general models narrowly. They also observe last year's zero-hire, high-ARR startups have mostly returned to normal hiring post-Series A, as rising customer expectations, not headcount efficiency, now seem to be AI tooling's dominant effect.
Dawn Song, a UC Berkeley computer science professor and co-director of the Berkeley Center on Responsible Decentralized Intelligence, ends up arguing that agentic AI cannot be trusted with critical actions until it can be evaluated in standardised, reproducible ways and, ideally, carry provable security guarantees, because the models at its core are powerful but not understood.
A broad career. Song moved from physics to computer science in grad school and says the MacArthur Fellowship gave her courage to broaden; after a startup acquisition she switched her group to deep learning before AlphaGo. She declined Hanselman's "talkers versus doers" framing of academics.
What agentic means. She says almost all researchers were surprised by the speed of progress, and nobody knows whether next-token prediction plus reinforcement learning will suffice.
Security in both directions. Her overview paper lays agents along dimensions of flexibility, each enlarging the attack surface. Agents are harder to secure against prompt injection and jailbreaks, and more dangerous when misused by attackers. Hanselman added that no one can hold the full stack in their head; Song went further: no one understands how LLMs work.
Jagged intelligence. On the stochastic-parrot debate she takes neither side, citing olympiad-level wins alongside simple errors, and benchmarks showing generalisation weakens as compositional complexity grows.
Teaching and evaluation. Her agentic AI MOOC, begun in fall 2024, has over 32,000 enrolled across three editions. Its competition, close to 1,000 teams last spring, now targets agent evaluation, which she says differs from model evaluation because the harness matters too, with over $1 million in prizes.
Pentland is a professor at MIT and Stanford HAI and the author of Shared Wisdom. He argues that AI fails because it treats people as logical individuals when they are social foragers. The remedy he ends with is AI that helps people think together rather than replacing them: personal agents loyal to their users, data cooperatives, and accountability through open audit trails.
The historical pattern. In Pentland's account, each AI generation failed on its model of society. Early optimal resource allocation, which he calls still the most common computation in the world, ran the Soviet economy until bad data and rare events undid it. Expert systems optimized for the average and centralized away local clinics and banks. Machine learning looks backward and breeds echo chambers. He tells the host that today's system could fall apart for the same reasons.
Social foragers. Citing Kahneman's estimate that about 95% of behaviour is learned from others, he argues that design must account for human context. Follower-driven social media amplifies extreme voices, and he says his group has shown this causes polarization. In his experiments, traders who watched other traders did better, and he reads 2008 as a failure to see tail risk.
Tools that advise. His loyal agents project, with Stanford Law School, addresses agents that serve the vendor, break on edge cases, act illegally or get lost in networks. AI buddies keep employees informed. Deliberation.io removes big voices and adds an AI mediator whose summaries people preferred two to one over human ones. In Washington, DC, residents wanted AI to defend them against the city. A citation-network buddy finds research blank spots surrounded by active researchers, where he says superstar papers come from; run on patents or legal citations, it flags commercial opportunity or churning law.
Data and governance. Pentland calls data a means of production and proposes data cooperatives modelled on agricultural banks and labor unions. For governance he proposes global, consistent accountability: open audit trails, protected by encryption or knowledge proofs, and long enough to catch slowly emerging harms. Standards bodies like UL would be the model, not pre-approval. He cites California's new safety-disclosure law and Stanford and Consumer Reports work on comparing labs.
Entrepreneurship and living labs. His programs have spun out more than 30 companies. His advice is to look where others don't, and to bet on fast-falling costs: he says the cost of AI at a given performance halves every 3.5 months, about 10x per year. Phone-based AI for rural health workers came before Wi-Fi did. Living labs, where communities agree to "live in the future", surface unanticipated effects. Digital twins let designs be tested on validated models of behaviour, protecting privacy and avoiding harm.
Hanselman and Mark's Ignite 2025 session, billed as teaching nothing, builds a deliberately useless chain from a 1975-era Altair to a 1,792-core cloud machine, and ends on one claim: the whole stack was vibe coded in a few days only because both hosts already understand the fundamentals underneath it.
One with the machine. Hanselman boots his kit-built Altair replica by keying a bootloader into its switches, runs an assembler Pong, and boots CP/M into Microsoft BASIC, which hangs. Mark notes that Azure servers still boot through UEFI, the same open-source firmware, being rewritten in Rust, used on Surface devices. Mark shows Compute! articles written at 16 in 1985 and a 1996 Sysinternals list whose one survivor is WinObj; the Sysinternals newsletter had 60,000 subscribers in 2006.
Zork in the cloud. Hanselman patches Zork to compile in Azure Cloud Shell, wraps it in a Docker container whose screen is fed to an LLM, and launches three and then 100 AI players from ACI standby pools, streaming scores over Event Hubs. He announces Zork I to III open-sourced live, with Xbox, Activision and Infocom, via a pull request to Historical Source.
Connecting the dots. Mark's Rust task manager on a 1,792-CPU Linux machine paints GIFs across its cores and runs an Altair emulator. Hanselman links the two: a webcam and NPU-run model read the physical Altair's LEDs at 28 frames a second, relayed by UDP, then WebSocket through ngrok, to Mark's emulator, "the least efficient way to do anything".
Vibe coding and fundamentals. Everything was vibe coded over four or five days of spare hours; open-sourcing Zork took longer because it involved humans. Mark argues reading code is easier than writing it, so experienced engineers get the boost and people who skip fundamentals do not. Hanselman frames modern Azure as declared desired behaviour, deployed with Bicep, and concedes the AI did nothing they could not have done themselves given time.
Marcus Fontoura, a technical fellow in Azure Core at Microsoft, ends up arguing in the book Human Agency in a Digital World that digital systems are human choices to be understood and shaped, with AI used to amplify human judgment on real problems and technologists taking a larger part in society's decisions.
Why the book. Fontoura's daughters' questions about AI, jobs and education prompted it; technical deep dives serve as illustrations so curious readers can reason about systems at a high level.
Evolution and fragility. Fontoura sees AI as an evolution from earlier machine translation and search, powered by data and compute. Asked whether companies over-optimise into fragility, Fontoura agreed but said software differs from bridges: it assumes unreliable parts and builds resiliency on top, citing Leslie Lamport.
Agency, juniors and teams. Fontoura would keep some activities human even against a perfect robot tutor, and shares Hanselman's concern that AI helps senior engineers more than juniors, calling for an intentional plan. Hanselman relayed Mark Russinovich's scepticism of huge context windows; Fontoura added that systems thinking is not yet encoded in AI, and that agents sharing one model make a more fragile team than heterogeneous humans.
Virtuous cycles. Productivity gains should grow the pie and be reinvested in innovation, not used to shrink teams.
Laws of fiction. Drawing on Harari, Fontoura separates physics from invented systems such as money and apps, arguing society can decide which it wants, as with social networks' effect on teenagers. Fontoura dismisses AI-domination fears as dated, wants AI aimed at climate, income distribution and transportation, and wants more computer scientists in politics.
James Hawkins, CEO and co-founder of PostHog, walks through how a company now valued at $1.4 billion — 160 employees, roughly 300,000 customers, and 15-16 products in production or development — was built out of repeated failure, and lands on a conviction that the next leap is building AI that replicates entire job functions inside a company, not just features.
Pivot hell, literally. Hawkins and his co-founder quit stable jobs to bootstrap out of coffee shops, working through roughly one idea every five to six weeks. Their first attempt, a sales-territory management tool, drew fifteen interested prospects but only one ever clicked a signup link — a lesson, Hawkins says, that salespeople are unreliable signal because they're friendly on calls but don't follow through. A second pivot, a developer technical-debt survey tool, also failed to land. What worked was born from the pivots' own byproduct: exhausted from re-setting-up product analytics after every idea change, they built a self-hosted, open-source version aimed at engineers rather than product managers, reasoning that technical users' stated interest tracks their actual behavior far more closely.
Discipline over waiting for inspiration. Hawkins is emphatic that the team didn't wait to feel excited about an idea — they built things to discover excitement, holding themselves to weekly output goals and meeting every interested prospect in person, sometimes traveling long distances. Charging real money early, even for a $50/month product, gave them an unambiguous signal: if people wouldn't pay despite the founders' full effort, the idea wasn't solving a real problem.
A launch, then a brutal raise. The open-source analytics tool launched on Hacker News about a month before YC demo day and became, by Hawkins's account, the year's most-upvoted dev tool post there, flipping demand from push to pull overnight. But the timing coincided with COVID's onset in March 2020, and VC interest evaporated within a week; PostHog ended up raising via $5,000 angel checks after meeting some 160 firms for its seed round — the hardest fundraise in the company's history.
From scarcity to conviction. By contrast, the recent $75 million round at a $1.4 billion valuation, making PostHog YC's newest unicorn, came together after conversations with essentially one investor who'd offered to lead preemptively. Hawkins traces the round's ambition to a growing personal conviction, formed partly by reading about AI lab founders' original reasoning, that AI's capability gains won't plateau — pushing PostHog to attempt "product autonomy": a tool that scans session recordings, analytics, error tracking, and LLM traces across all its products and opens pull requests fixing issues automatically, aiming to build an AI equivalent of a role, product management, that already has proven market fit.
Standing out on purpose. Hawkins also credits deliberately "remarkable," polarizing marketing — absurdist billboards unconcerned with direct conversion and a sprawling, unconventional website documenting all 16 products, a developer-focused jobs board, and a public company handbook — with building the trust and attention a crowded, increasingly AI-saturated market demands.
Li, a senior data scientist at Netflix who previously worked at Spotify and DoorDash, argues that practical data science at scale rewards the unglamorous: repeatable infrastructure, interpretable forecasting models, and expertise in the manual work before any automation. He makes the same case for careers. What got him into Netflix was not a tactic but a rare combination of skills, ads plus forecasting, at the moment Netflix needed it.
Scale versus zero-to-one. The host frames Li's work as Spotify's ad business of over a hundred million dollars a year and Netflix's billion-dollar supply-and-demand forecasting. Li says zero-to-one work, like starting Spotify's podcast ads, is hacky and fast. Scale demands naming conventions, reusable frameworks, and model auditing and alerting. Li explains that Spotify bought Megaphone, alongside big exclusive deals and Anchor for small creators, to own the whole podcast value chain.
Forecasting basics and pitfalls. Li describes most time-series models as linear regression adapted for autocorrelation: ARMA, exponential smoothing, and Prophet, which decomposes a series into trend, seasonality and error. The technical rookie error is a random train-test split. He uses expanding-window or sliding-window cross-validation instead. The bigger pitfall is reaching for LSTMs and transformers when stakeholders need to know why a forecast says what it does. The host says that in his 2016–2019 deep learning courses, finance students never beat simple statistical methods with deep nets. Li says the M forecasting competitions were long won by univariate methods, and that hybrid ensembles of tree-based, deep learning and univariate models now perform best. The edge from complex models may matter to a quant chasing 1.1% more accuracy, he adds, but rarely at work.
Foundation models. His team tested time-series foundation models including Chronos and TimesFM, and none beat its simple approaches. He would switch for significantly better accuracy and a business case, but explainability remains a barrier. He splits forecasting into operational uses, where accuracy rules and deep learning can fit, and strategic uses such as FP&A, where assumptions must be visible. He credits this split to a forecasting course he took.
Getting hired. The briefcase technique, a detailed pain-points doc for the hiring manager, got him into DoorDash. The host worries that such docs now look like AI spam, citing the 30-page outreach PDFs he receives weekly. Li answers that genuine specificity still lands. Netflix rejected him multiple times before its new ads business needed his ads-plus-forecasting mix. His advice is to persist and to build a niche at an overlap of skills.
Dating and AI. His former startup yourmove.ai, which the host says has over 10,000 users, began in the GPT-2 era. It grew from Li's 2019 model that auto-swiped profiles based on his own swipes. He says its suggestions helped but couldn't be fully automated, and that AI-generated photos still read as fake and count against you. He sees a shift toward in-person meeting and notes he met his girlfriend through skiing; she filtered dating apps for men 6 feet and over, and he is 5'9". The host relays news that Match Group's share price fell as post-pandemic dating-app activity dropped, with apps betting on AI.
Automation and agents. His failed ad-creative automation taught him to master a workflow manually first. He says he spent about $200 of API credits learning to prompt image models. He softens his "agents are just Python scripts" tweet: let the LLM decide whether to trigger, and keep execution deterministic.
Tools and mental models. AI has made him favour monorepos, because coding agents find context in one place more easily. His stack at the time of recording: Wispr Flow for dictating rich prompts (he writes specs by hand to think them through), Claude Code, Cursor's new Composer 1 model, Gemini's Nano Banana for images that keep his face intact, and v0 for front ends. He says the best tool for a job changes "week to week". He borrows mental models from Charlie Munger, including asymmetric upside, such as going to parties with interesting people, and activation energy.