We can't find the internet
Attempting to reconnect
Something went wrong!
Hang in there while we get back on track
1. Funding & Deals
Conveo AI raised a $50M Series A led by DST Global. Conveo (YC S24) says its AI interviewer conducts in-depth video conversations with thousands of consumers, allowing brands to compress research cycles from months to days; YC says more than 400 enterprises use it, including 50+ Fortune 500 brands such as Google, Unilever, and Canva. This is a clean application-layer bet: turning the depth of focus groups into a scalable workflow without requiring a new foundation model. The diligence question is whether interview quality, enterprise distribution, and accumulated workflow data create durable advantages rather than a cheaper interface to general models.
2. Emerging Teams
A personal-trainer SaaS team reports unusually fast early revenue, with domain knowledge doing as much work as the product. The team says it launched eight months ago after ten months of building and is at $450,000 ARR. It reports more than seven years of domain experience, prior work with major coaches, and a deliberate focus on higher-performing coaches rather than a generic market. Competitor-specific Instagram outreach produced 20–25% of its first 50 customers; it now reports $15,000/month in advertising spend, roughly $200 ARPU, 1.2% monthly churn, and about 10x ROAS, while acknowledging a four-month payback period and weak UTM attribution. The signal is the combination of narrow ICP, operator credibility, and distribution; the metrics remain founder-reported and should not be underwritten without cohort and payback verification.
Dreamwork is turning job-search traction into a high-externality automation bet. The founder reports growth from 2,000 to 7,500 authenticated users in two months, roughly 90 paying users, and users being hired through the product. Its planned Autopilot selects matches at 70% or above, tailors a resume and cover letter, and applies on the employer’s ATS rather than through a back channel. The founder acknowledges the obvious downside—that mass application could flood recruiters and degrade the ecosystem—while arguing that laid-off workers need representation and saying guardrails are in place. For investors, this is a live test of whether labor-market agents optimize for application throughput or defensible quality; hiring outcomes, stale-listing rates, and factual-error rates matter more than user growth alone.
Payelle shows a different early-stage wedge: pre-transaction card selection rather than payment routing. After 18 months, the team says it moved from a physical-card concept to mobile wallets, built the technology, filed IP, launched, and entered conversations with large companies and financial institutions, while still describing distribution as unresolved. Its technical claim is to map a merchant’s business category to the correct MCC and surface the best card before the tap; the founder calls the system a work in progress but “exceptionally accurate.” Amex reportedly contacted the team while it was still in stealth after seeing the founder’s thesis, an encouraging institutional signal that is not the same as a partnership. Apple Wallet surfacing is live, while Android still requires additional partnerships.
3. AI & Tech Breakthroughs
Patronus Ark makes selective-depth inference a concrete AI-security architecture. The product detects prompt injections, dangerous tool calls, sensitive documents, and PII on the endpoint, targeting ordinary business laptops without GPUs. Its three-stage stack combines heuristics, cheap classifiers, and a full mmBERT model; the cheap and deep stages share representations, and a learned gate decides whether deeper inference is needed. Patronus says only 16–24% of cases enter the full path, eliminating roughly 76–84% of full-transformer executions. Five of eight evaluated pipelines remain within one F1 percentage point of always running the full model, while the current desktop application reportedly peaks at about 750 MB and 500 ms median analysis latency. The caveat is material: three pipelines still lag, and further layer-skipping results require ablation. The broader investment thesis is selective inference for the large volumes of tool output and retrieved content that agent-security systems will need to screen.
Agent harnesses are emerging as an independent performance and cost layer. FrontierHarness Eval holds the model, tasks, and runtime constant across 360 runs and 2 billion tokens, yet reports pass rates ranging from 50% to 67% and cost per pass from $1.05 to $18.34. Martin Casado’s description of Exo makes the design implication explicit: the model sees the entire harness code, running code, and logs, with the ability to upgrade the runtime dynamically—not just the prompt. The market consequence is that model selection alone will not explain agent outcomes; execution loops, state, tooling, and evaluation become part of the product moat.
World Labs’ current signal is a claimed real-to-sim-to-real robotics workflow. Justin Johnson describes taking five phone photos of a space, reconstructing it in Atlas, specifying a task in natural language, having an agent build the simulation, and reinforcement-learning fine-tuning a robotics foundation model for that environment in roughly five minutes. If repeatable, this would attack one of physical AI’s hardest bottlenecks—environment-specific data collection—by turning casual observations into training environments. The evidence here is a founder explanation of a workflow, not deployment or robot-performance results.
Perplexity open-sourced Lily, making local inference a reusable product component. Lily is the local inference engine behind hybrid compute in Perplexity Computer, specialized for Qwen3.6-35B-A3B on Apple Silicon so on-device computation does not bottleneck Computer tasks. The important shift is from local execution as a privacy slogan to local execution as an engineered distribution asset that can be embedded in agent products.
4. Market Signals
The bullish company-formation thesis is colliding with a physical infrastructure bill. At the G20, Sam Altman argued that work once expected from a startup during a three-month accelerator is now “probably doable in like 17 minutes with Codex,” enabling faster testing, building, and customer feedback; he also expects persistent agents to act as virtual collaborators. His adoption case is explicitly infrastructure-heavy: countries will build or rent data centers, and even major efficiency gains will not remove the need for much more capacity if AI is to remain abundant and inexpensive. The caution is equally explicit: cyber- and biosecurity failures could set adoption back, while concentrated compute could worsen inequality.
The hardware market is already showing the scale of that buildout, but not necessarily equivalent value capture. An AI-infrastructure analysis cites Dell’s $16.1B of AI-optimized server revenue in Q1 FY27, $24.4B in AI orders, $51.3B in ending backlog, and roughly $60B in expected full-year AI-server revenue. It also notes that GPUs, HBM, interconnect, and cooling flow through Dell systems while the highest-rent silicon layers sit elsewhere; the underwriting question is whether deployments let Dell move into storage, networking, orchestration, services, and financing.
SaaS is being pulled into existing agents, but access alone is not enough. A current founder discussion argues that buyers increasingly want to stay in Claude Code, Codex, or a terminal and invoke products through APIs, MCP, or execution hooks rather than open another walled-garden UI. Glean reports that users spend about half of a five-hour AI day building context and that its Claude/Cursor/Codex MCP path is growing faster than its own UI; it distinguishes runtime retrieval from the offline work of mapping people, teams, projects, and acquired-company history. When agents write to systems of record, identity and a trace rich enough to investigate errors become product requirements. The investable layer is therefore API access plus durable organizational context and accountability, not simply another assistant sidebar.
Capacity and trust are both becoming adoption bottlenecks. One startup says OpenAI and Anthropic endpoint TPM limits across AWS, GCP, and Azure are hit during usage spikes, preventing new pilots and reliable SLAs; repeated quota requests reportedly went unanswered for months. In recruiting, Dreamwork’s auto-apply approach is countered by Match Moth, which explicitly refuses to apply for users, shows its scoring rationale, crawls company career pages, rechecks listings nightly, labels ghost jobs, and locks employers, dates, and numbers against invention. The emerging product split is throughput versus user control and auditability—an important design choice for any agent acting in a consequential workflow.
5. Worth Your Time
- Watch — Sam Altman and Howard Lutnick at the G20 Innovation Ministerial. The useful part is not the “AI is inevitable” rhetoric but the combination of a concrete 17-minute company-formation claim, a compute-abundance thesis, and explicit cyber, biosecurity, and concentration risks.
- Watch — The Operations Startup Managing $6B of GMV | Lightwork. Zach Pang’s account of Seal is a practical case for building accountable AI agencies rather than selling customers another agent toolkit: the platform runs support, resale, and compliance for thousands of brands and marketplaces, while its founder links resolution limits to missing operational context.
- Read — The CPOs of Harvey, Glean and Rubrik on What It Actually Takes to Ship a Category-Winning Agent. It is a compact enterprise deployment checklist: start with non-destructive workflows and approval gates, measure the hidden cost of context-building, and require identity and traceability before agents write into systems of record.
Conveo raised a $50 million Series A led by DST Global, according to CEO Dieter De Mesmaeker.
The supplied source does not identify any participating investors beyond the lead and does not state an investment thesis or use of proceeds. It only describes Conveo as helping big brands understand market trends.
- Moderna and Merck reported that an interim Phase 3 trial of an individualized mRNA cancer vaccine in melanoma met its recurrence-free-survival primary endpoint and distant-metastasis-free-survival secondary endpoint; the CEO emphasized that this was a first interim analysis, not the completed study, with data planned for a major oncology conference. He characterized the result as the first cancer-vaccine success after more than 20 years and over 1,000 failed clinical trials. The company hoped to file with regulators and make the treatment available in 2027.
- The personalization workflow compares tumor and healthy-cell DNA nucleotide by nucleotide, uses an algorithm to select the 34 most relevant mutations, and produces a patient-specific mRNA molecule in roughly 30 days; cited Phase 2 data indicated that about 90% of antigens differ from one patient to another.
- The current biopsy-to-hospital cycle is about 42 days, but unlike ex-vivo cell therapies it requires only sequencing information: the product is made through synthetic, enzymatic manufacturing in comparatively small reactors. The company said its facility could produce tens of thousands of doses, while pricing and cost of goods had not yet been disclosed.
- The proposed regulatory model is a process BLA rather than approval of every individualized dose; the company has engaged the FDA for years and must demonstrate that identical tumor-and-blood inputs reliably produce identical patient outputs.
- The platform is expanding beyond melanoma into Phase 3 lung cancer, Phase 2 kidney and bladder cancer, and a Phase 3 monotherapy study in Stage 1 lung cancer; pancreatic and gastric cancer are higher-risk targets because prior checkpoint studies were negative. The CEO also plans to use AI to mine clinical data, especially the roughly 20% of patients who do not respond, to improve the algorithm. Beyond oncology, the company is targeting a pivotal study in a rare pediatric liver disease and exploring both shared and individualized mRNA approaches for autoimmune disease, with the individualized approach still in the lab.
- Scaling thesis: Sam Altman attributes the AI inflection to deep learning beginning to improve with more compute in 2012. OpenAI’s early work moved from robot-hand and video-game experiments into unsupervised learning and GPT models; he describes GPT-3 as crossing a usability threshold and GPT-4 as the point when internal skeptics recognized the approach could go much further.
- Agent-native company formation: Altman says coding agents and white-collar-work agents accelerated in 2025, with persistent agents that act as continuously available virtual collaborators likely to follow. Drawing on his startup-accelerator experience, he says work previously expected from a startup over three months may now be achievable in roughly 17 minutes with coding-agent tooling, pointing to sharply compressed early-stage build cycles.
- Application-layer opportunity: OpenAI’s stated role is to provide a general AI engine or commodity, while entrepreneurs, governments, and companies supply the domain context, questions, and workflows. This supports an investment thesis around context-rich applications and vertical execution layers built on frontier models.
- Infrastructure intensity: Altman expects AI usage to grow so substantially that, even with more efficient chips and algorithms, the ecosystem will need much more infrastructure; countries may choose to build data centers or rent capacity. He argues that keeping AI abundant and inexpensive is necessary to distribute its benefits broadly rather than concentrating access among wealthy users.
- Adoption signal and risks: Altman calls national AI adoption effectively “non-negotiable” because of the expected economic and quality-of-life benefits across businesses, education, healthcare, and services. He simultaneously flags urgent cybersecurity and biosecurity challenges and warns that concentration of AI power or compute could worsen inequality and delay adoption.
- Agentic and vertical AI are a major startup layer: Large language models become agents when surrounded by retrieval, working memory, tools, and collaboration; the same architecture can operate digital or physical systems, including manufacturing robots, logistics vehicles, surgical robots, and drug-discovery labs. Anthropic’s Tom Brown said the company focuses on raw models while startups adapt them with domain knowledge and tools for legal, biology, and mechanical-engineering workflows, describing this surrounding opportunity as much larger than the model layer itself.
- Compute infrastructure remains a bottleneck and investment theme: Brown identified data centers and compute as “very clearly the bottleneck,” citing shortages of power and labor; land permitting and construction labor are key prerequisites, while host countries can gain investment, jobs, and tax revenue.
- Technical pedigree is concentrated around scaling-law research: Brown described himself as an engineer, an early OpenAI employee among its first 20, and Anthropic’s chief compute officer; he said Anthropic’s co-founders developed the 2019 scaling-laws research that linked compute, data, and model size to intelligence and helped motivate Anthropic’s founding on a sub-10-year AI timeline.
- Model economics are moving rapidly: Brown expects capability gains to continue at roughly the current pace and said access to a model comparable to the best model of the prior year is now roughly 20 times cheaper.
- IP and regulatory treatment remains a commercialization risk: The policy discussion called for protecting patents, copyrights, and trade secrets while preserving AI competition and commercialization, with fair-use and fair-dealing rules expected to depend on jurisdiction and circumstances.
- Technical scaling and agent trajectory: Sam Altman attributes the AI inflection to deep learning beginning to work and improve with more compute in 2012, while OpenAI’s scaling-law findings motivated pursuing much larger compute budgets. He says the GPT-4 period was when even internal skeptics recognized the approach could go far, with current systems capable of discovering knowledge, doing science, and producing complex software.
- AI-native company formation: Altman says founders who lack capital, experts, or operational resources can now use AI to develop ideas, manage complex supply chains, and pursue new science; he identifies coding and white-collar-work agents as a major 2025 acceleration and expects persistent agents to function as virtual collaborators. Drawing on his accelerator experience, he estimates that work previously expected from a startup over three months may now be achievable in about 17 minutes with Codex, enabling faster testing, building, and customer feedback.
- Application-layer opportunity: OpenAI positions its role as providing a general AI engine or commodity, while entrepreneurs, companies, and governments supply the context, questions, and domain problems. OpenAI also says it wants to operate as a platform and does not want to be the only company capturing the value, pointing to room for early-stage companies building domain-specific applications on top of general models.
- Compute and deployment infrastructure: Altman argues AI adoption is effectively non-negotiable and predicts that products and services will become broadly intelligent and AI-enabled. He says even major efficiency gains, better chips, and smarter algorithms will still require much more infrastructure to make AI abundant and inexpensive, while countries may choose to build or rent data centers.
- Cautionary signals: He flags cybersecurity as an urgent challenge and cites biosecurity and future risks that, if mishandled, could substantially slow AI adoption. He also warns that concentrated power or scarce compute could turn AI into a privilege for wealthy users and worsen inequality.
- Founder signal: Sam Altman says he moved from studying AI into entrepreneurship, later ran a startup accelerator, and started OpenAI at the end of 2015 after recognizing the potential of deep learning and compute scaling.
- Startup formation and agent productivity: Altman expects AI to drive a major expansion in entrepreneurship and small-business creation by giving founders access to expertise and execution capacity they previously lacked. He says coding agents and white-collar agents accelerated in 2025, with persistent agents acting as virtual collaborators likely to follow; work previously expected from a startup over a three-month accelerator is, in his estimate, now achievable in about 17 minutes with AI coding agents.
- Application-layer thesis: Altman says OpenAI should provide an AI “engine” or commodity, while entrepreneurs, companies, and governments supply the context, questions, and domain-specific problems; this points to continued opportunity for contextualized products and workflows built around general AI infrastructure.
- Infrastructure demand and adoption risk: Altman argues that AI usage is scaling rapidly enough that, even with more efficient chips and smarter algorithms, the industry will need much more infrastructure to keep intelligence inexpensive and abundant. He warns that cybersecurity and biological-security failures could set adoption back, while concentrated access to compute could concentrate power and worsen inequality.
- Sam Altman says he studied AI, moved into entrepreneurship when early AI progress stalled, recognized the 2012 deep-learning/compute inflection, and started OpenAI at the end of 2015; he also had prior experience running a startup accelerator.
- Altman identifies compute scaling as the core technical thesis behind the current AI wave: deep learning began working in 2012, capability improved with more compute, and the GPT-4 period convinced even internal skeptics that the approach would go far; he describes newer systems as able to discover knowledge, do science, and create complex software.
- AI is lowering barriers to company creation: Altman says people without enough capital, resources, or expert hires can use AI to pursue ideas, develop science, and manage complex supply chains; he cites a laundromat operator using ChatGPT for marketing, legal-contract advice, and vendor relationships.
- Agentic software is a near-term theme: Altman says 2025 saw acceleration in coding agents and white-collar-work agents, and expects persistent, long-term agents to function as continuously helpful virtual collaborators.
- AI-native development is compressing experimentation: work Altman expected a startup to accomplish during a three-month accelerator is now “probably doable in like 17 minutes with Codex,” allowing founders to test ideas, build products, and get customer feedback much faster.
- Altman urges countries to treat AI adoption as non-negotiable, comparing it with electricity, while saying governments may build data centers or rent them; this makes compute access and national AI deployment important infrastructure themes.
- His positioning is pro-adoption but not unqualified: he describes AI as bringing both excitement and anxiety, says systems need constraints against catastrophic risks, and emphasizes respect for national sovereignty and diverse uses.
- Agentic software is compressing startup execution: Sam Altman says coding and white-collar work agents accelerated in 2025, with persistent long-term agents expected next; he estimates work that once took a startup accelerator three months can now be done in roughly 17 minutes with Codex, pointing to much faster prototyping, customer feedback, and company formation.
- Contextual application companies remain an explicit opportunity: Altman describes OpenAI’s role as providing a general AI engine, while entrepreneurs and institutions supply the questions, domain context, and problems to solve; he positions OpenAI as a platform that wants the broader ecosystem—not just OpenAI—to capture value.
- Compute abundance is a major infrastructure thesis, with adoption risks: Altman argues that rapidly rising AI usage will require substantially more infrastructure, improved chips, and more efficient algorithms to keep intelligence inexpensive and broadly available; he specifically flags cybersecurity, biosecurity, and concentration of compute or power as risks that could materially slow adoption.
- Agent-driven startup formation: Sam Altman said coding agents and white-collar-work agents accelerated in 2025, with long-term persistent agents that act as virtual collaborators as the next step. He argued that work a startup was expected to accomplish during a three-month accelerator could now be done in “like 17 minutes with Codex,” enabling faster idea testing, customer feedback, and iteration.
- Application and context layer: OpenAI’s stated role is to provide an AI engine or commodity, while people and companies supply the questions, use cases, and domain context; this points to opportunity for startups that encode specialized context and workflows on top of general AI rather than only building models.
- Infrastructure demand and adoption risk: Altman said efficiency gains, better chips, and smarter algorithms will still require much more infrastructure as AI usage scales, and argued that AI must become abundant and inexpensive to avoid compute concentration and inequality. He separately warned that cybersecurity and other safety failures could materially slow adoption and “set this technology back a great deal.”
- AI is compressing early-stage company-building cycles. Sam Altman said coding agents and white-collar-work agents accelerated in 2025 and expects persistent agents that act as virtual collaborators next. He said work that once filled a three-month startup accelerator could now be done in “like 17 minutes” with Codex, enabling faster idea testing, product building, and customer feedback.
- Model providers may supply infrastructure while application builders supply context. Altman described OpenAI’s role as providing an AI “engine” or commodity, with people and companies responsible for the questions, domain context, and applications. He said countries may build or rent data centers, but demand will require substantially more infrastructure even as chips and algorithms become more efficient if AI is to remain abundant and inexpensive.
- The adoption thesis carries significant safety and concentration risks. Altman warned that cybersecurity and biosecurity failures could set AI adoption back substantially, while concentrated compute and power could worsen inequality if AI is not made abundant. He identified GPT-4 as the point when internal skeptics recognized the approach’s potential and described newer systems as capable of discovering knowledge, doing science, and producing complex software.
- Founding pedigree: Sam Altman says he studied AI, pursued entrepreneurship, and started OpenAI at the end of 2015, with work beginning in January 2016, after recognizing that deep learning improved with more compute.
- AI-native company formation: Altman says 2025 brought acceleration in coding agents and white-collar-work agents, with persistent long-term agents potentially emerging next. He argues that work a startup was expected to accomplish during a three-month accelerator can now be done in roughly 17 minutes with Codex, enabling faster idea testing, product building, and customer feedback.
- Application and infrastructure thesis: Altman frames OpenAI as an AI engine or commodity, while entrepreneurs and operators provide the domain context, questions, and problems that create differentiated applications. He expects much more infrastructure to be needed despite efficiency gains in chips and algorithms, to keep AI abundant and inexpensive; he gave an explicitly approximate comparison of token use rising from roughly 100,000 per month for a leading user in early 2020 to hundreds of billions per month for a leading user he knew by mid-2026. He also flags cybersecurity, biosecurity, and concentration of compute as risks that could delay adoption or worsen inequality.
- Founder pedigree: Sam Altman says he studied AI, moved into entrepreneurship, ran a startup accelerator, and started OpenAI at the end of 2015, with work beginning in January 2016.
- AI-native company formation: Altman predicts an unprecedented boom in entrepreneurship and small-business creation as AI gives founders access to expertise, capital-like capabilities, science development, and complex supply-chain management. He says coding agents and white-collar-work agents accelerated in 2025, with persistent agents acting as continuous virtual collaborators as the next product paradigm; work once expected during a three-month accelerator may now be achievable in roughly 17 minutes with an AI coding agent.
- Application-layer opportunity: Altman describes OpenAI’s role as providing an AI engine or commodity, while companies and entrepreneurs supply the domain questions, ideas, and organizational context; this suggests room for startups building contextualized workflows and products on top of general AI systems.
- Infrastructure investment theme: Altman expects major efficiency gains from better chips and algorithms but still says much more infrastructure will be required to make AI abundant and inexpensive; countries may choose to build their own data centers or rent capacity.
- Cautionary flags: He identifies urgent cybersecurity and emerging biosecurity challenges as risks that could significantly set AI adoption back, while concentration of power and scarce compute could worsen inequality if AI is not made abundant.
- AI agents are sharply lowering startup execution costs. Sam Altman says coding and white-collar agents accelerated in 2025, with persistent agents that act as virtual collaborators likely to follow; he claims work a startup accelerator expected founders to complete in three months can now be done in “like 17 minutes” with Codex, enabling much faster idea testing, product building, and customer feedback.
- The application and context layer remains open around foundation-model platforms. Altman describes OpenAI’s role as providing an AI “engine” or commodity, while businesses and entrepreneurs determine the questions, ideas, and domain or societal context to apply it to. OpenAI also says it does not want to be the only company capturing the value and aims to operate as a platform and service to the broader economy.
- Compute and AI infrastructure remain a major investment theme. Altman expects continued gains from more efficient systems, better chips, and smarter algorithms, but says substantially more infrastructure—including data centers that countries may build or rent—will still be required to make AI abundant and inexpensive.
- Key cautionary signals are cybersecurity, biosecurity, and concentration of compute and power. Altman warns that urgent cybersecurity failures could materially set back AI adoption, while scarce or concentrated AI infrastructure could worsen inequality; he argues that abundance is necessary for AI to function as an equalizing force.
- Founder pedigree signal: Altman says he studied AI, moved into entrepreneurship after early AI techniques failed to work, then founded OpenAI after recognizing the 2012 deep-learning and compute-scaling inflection; OpenAI was announced at the end of 2015 and began work in January 2016.
- ApplyBoard addressed the cross-border university application gap for international students by simplifying applications to U.S. universities; its school-side monetization was a $5,000–$10,000 affiliate bounty per admitted student.
- The episode presents ApplyBoard as a significant early angel miss: the host planned a $25,000 investment when he loosely estimated the company’s valuation at roughly $5 million–$10 million, but failed to follow through. He later estimated more than $1 billion in annual revenue and described the company as multi-billion-dollar, while explicitly noting that he did not know the exact figures.
- Agent-enabled startup formation: Sam Altman says coding and white-collar agents accelerated in 2025, with persistent agents that act as virtual collaborators potentially coming next. He estimates that work expected from a startup during a three-month accelerator may now be achievable in roughly 17 minutes with Codex, enabling much faster idea testing, product building, and customer feedback.
- Platform and application-layer opportunity: Altman describes OpenAI’s role as providing a general AI engine or commodity, while entrepreneurs and companies supply the context, questions, and applications; he says OpenAI does not want to capture all the value or be the only company.
- AI infrastructure demand: Altman says scaling laws link additional compute to greater model capability and argues that, despite continued efficiency gains in chips and algorithms, much more infrastructure will be needed to make AI abundant and inexpensive rather than a scarce, highly priced commodity.
- Adoption risks: He flags cybersecurity as an urgent challenge and also points to biosecurity and excessive concentration of compute and power; failures in these areas could substantially slow AI adoption.
- Founder pedigree: Altman says he studied AI, moved into entrepreneurship, and later ran a startup accelerator before starting OpenAI, which was announced at the end of 2015 with work beginning in January 2016.
- Sam Altman argues AI is lowering the capital, expertise, and operational barriers to starting a company, potentially driving an unprecedented boom in entrepreneurship and small-business creation; he cites a laundromat operator using ChatGPT for marketing, legal-contract advice, and vendor management.
- Agentic software is emerging as a major product paradigm: Altman describes acceleration in coding agents and white-collar-work agents in 2025, with long-term persistent agents that function as virtual collaborators as the next expected step.
- The opportunity may concentrate in the application and context layer rather than only foundation models: OpenAI intends to provide a general AI “engine,” while entrepreneurs and organizations determine the domain-specific questions, workflows, and societal context; Altman also says OpenAI does not want to capture all economic value.
- AI adoption is becoming a national competitiveness imperative, with governments choosing among different regulatory approaches and infrastructure strategies, including building domestic data centers or renting capacity elsewhere.
- AI is being positioned as a capability amplifier for individual professionals: the video cites front-end developers using AI for back-end code, marketers building web-crawling dashboards, recruiters executing technical workflows, and CFOs writing Python automations; it links this expansion of individual scope to rising demand for high-agency workers. AI outputs remain inconsistent, making human context, business judgment, and domain expertise important complements.
- A cautionary product signal is cognitive offloading: the video claims AI can improve students’ immediate assignment performance while weakening later retention, and recounts Andrew Ng forgetting how AI-generated project components worked six months later.
- Agent harnesses are becoming a distinct optimization layer: FrontierHarness Eval compares Pi, Exo, Claude Code, Codex, DeepSeek Harness, and four others using the same model, tasks, and runtime across 360 runs and 2 billion tokens; reported pass rates range from 50% to 67% and cost per pass from $1.05 to $18.34. Martin Casado highlighted Exo as the best price/performance result in this comparison.
- Exo’s technical differentiation is full-runtime model access: its design exposes the entire harness code, running code, and logs to the model—not just the prompt—with the ability to upgrade dynamically for self-improvement; Casado describes the approach as “maximally bitter lesson aligned.”
- Zach Pang founded Seal in 2019 after working at Orbital Insight; he studied philosophy and physics, learned to code there, and brought algorithmic prediction experience from satellite-data applications for hedge funds and insurers into e-commerce. Seal’s initial wedge was underwriting returns and refunds from pre-purchase behavior, targeting a market Pang contrasted with payment chargebacks: returns were about 15–25% of GMV versus chargebacks below 1%.
- Seal’s pre-LLM prediction system used deep learning, convolutional neural networks for computer vision, and roughly 200 pre-purchase signals, including product-page dwell time, purchase preferences, and whether apparel imagery showed a human model. Pang says the model initially produced deeply negative contribution margins; scale and large volumes of labeled orders were required to make the underwriting work, while marketplace breadth accelerated learning across categories.
- Seal has expanded from returns assurance into an AI operations platform whose workforce handles support, resale, and compliance for thousands of brands and marketplaces, including TikTok Shop and Poshmark. Its product thesis is to build AI “agencies” accountable for end-to-end KPIs rather than agent tools that leave customers responsible for prompts, testing, and calibration; Pang attributes the roughly 50–60% resolution ceiling of many agentic support products to missing unstructured operational context. The company is rolling out a subscription for Shopify startups and reports seeing one- or two-person teams handle volumes that would typically require 10–20 people, pointing to an emerging AI-enabled micro-enterprise and commerce-operations theme.
- Pangram raised $9M recently and partnered with Substack; Substack’s tool lets readers scan posts, notes, replies, and comments over 100 characters for a human-versus-AI estimate. The interview does not identify the round stage or lead investor. Quora also still uses Pangram to deter fully AI-generated answers, while the company is pursuing broader platform integrations and positioning itself as core internet infrastructure for distinguishing human from automated content.
- Co-founder and CEO Max Spiro says Pangram was started by machine-learning researchers to detect AI writing, evolving from binary detection into estimation of the degree of AI assistance. Its model uses licensed pre-2022 human writing, matched AI-generated examples, a 50/50 human/AI training set, and prompts representing AI editing such as “make it more detailed” or “polish my writing.”
- Pangram is extending from text into image detection as AI images become difficult to distinguish from photographs, with potential applications in misinformation and AI-enabled fraud such as fabricated receipts, food-contamination claims, and fake marketplace listings. Reliability is a material adoption and liability caveat: the CEO reports a text false-positive rate of about 1 in 10,000 and an image false-positive rate of about 1 in 1,000.