100 AI people
to watch.
The bets
behind them.

Reasoning needs an auditable result

According to Wei’s research page, Wei identifies his OpenAI research role and lists his participation in the CICERO team.

The business question concerns the gap between an impressive answer and an argument that can survive external checking. Reliable reasoning could reduce expensive expert review, but the review burden is part of the product economics.

The people behind reinforcement learning

According to OpenAI’s o1 contribution record, OpenAI credits Jerry Tworek with overall leadership within its reasoning-research section.

That is a more precise starting point than a talent-market headline. I want to understand whether the training approaches associated with reasoning systems become repeatable tools for useful work.

Making reasoning usable

According to Wei’s own research page, Wei says he works at Meta Superintelligence Labs and lists chain-of-thought and instruction-tuning work.

I am interested in the cost of turning a promising research technique into a dependable user experience. More reasoning is valuable only when the extra computation improves the result enough to justify its cost.

Attribution before the talent story

According to OpenAI’s o1 contribution record, Hongyu Ren appears among the foundational contributors to o1 reasoning research.

I would follow subsequent attributable research rather than use recruiting reports as a proxy for technical progress. The economic question is whether the work improves the cost or reliability of a reasoning system.

Where the reward signal comes from

According to Douglas and Bricken’s Claude discussion, Sholto Douglas discusses reinforcement learning and the difficulty of specifying useful rewards outside games.

I am watching that design problem because businesses need agents rewarded for finishing the right task, not merely producing persuasive output. A cheap automated reward that misses the real objective can create expensive errors.

Reasoning that a checker can inspect

According to Schrittwieser’s AlphaProof account, Schrittwieser describes the AlphaProof system and links its published mathematical-reasoning paper.

Formal checking offers a clearer success condition than a convincing paragraph. I am watching whether that discipline spreads to useful software or scientific reasoning, while keeping formal benchmarks separate from commercial readiness.

Reverse-engineering the model

According to Nanda’s account of his work, Nanda says he leads DeepMind’s mechanistic interpretability team and identifies TransformerLens among his projects.

I see interpretability as an attempt to lower the cost of understanding failures that ordinary testing misses. The interesting business consequence would be better deployment decisions, rather than an explanation attached after something goes wrong.

Training environments without building every world

According to Parker-Holder’s research page, Parker-Holder describes work on world models and points to Genie as an example.

Synthetic environments could change the economics of training agents, but a convincing generated scene is not automatically a faithful simulator. I am watching the link between simulated experience and measurable performance elsewhere.

Separate self-improvement from the slogan

According to Weng’s dated research notes, Weng’s July 2026 notes discuss recursive self-improvement; her blog also examines test-time computation.

Her useful contribution is making the technical distinctions inspectable. I would separate better outputs, better training pipelines and a system modifying itself, because their evidence requirements and business implications are different.

Announcing a model and delivering it are different events

According to Reflection’s Beam announcement, the October 5 announcement promised the model’s weights and supporting artifacts later that month.

My focus is the sequence from announcement to accessible artifacts to independent evaluation. A model’s economic usefulness depends on the cost of serving useful outputs, not simply the size of its parameter count.

Further records: NVIDIA's coalition announcement

Efficiency is also an engineering problem

According to The MiMo-V2-Flash technical report, Fuli Luo appears on the report’s author list for a model using a mixture-of-experts and hybrid-attention architecture.

The question for 2027 is whether the serving system captures the architecture’s intended savings under actual workloads. A theoretical reduction is useful only if engineering turns it into reliable, cheaper service.

Learning from feedback without retraining every time

According to The Reflexion paper, Shunyu Yao is a co-author of research that uses linguistic feedback and episodic memory to improve an agent’s subsequent attempts.

That raises a practical question about where improvement should happen: in the model weights, the surrounding software or the task history. I would compare those choices by reliability, maintenance effort and cost.

Open models need a usable deployment path

According to The Qwen2 technical report, Junyang Lin is listed among the authors of the Qwen2 model-family report.

The business question concerns the combination of model capability, accessible artifacts and practical deployment. The paper documents a research contribution; it does not establish subsequent financing or current employment.

Agentic models need workflow economics

According to The Kimi K2 technical report, Zhilin Yang appears on the author list of the Kimi K2 report, which describes an agentic model and released checkpoints.

My focus is the cost of completing an entire workflow, including retries and tool use. A good benchmark result can be a starting point, but it does not establish a low cost per successfully completed business task.

The robot has to learn outside the demonstration

According to Finn’s Stanford page, Finn identifies her Stanford appointment, Physical Intelligence co-founding role and research on learning through robotic interaction.

The business question concerns the amount of new data and human intervention needed when a robot encounters an unfamiliar setting. That determines whether a general-purpose system can scale beyond a carefully prepared environment.

Generative models as environments for action

According to Yang’s research page, Sherry Yang identifies her NYU and DeepMind roles and a focus on reinforcement learning and generative modeling.

I am interested in how models support decision-making, rather than just generate plausible outputs. The economic test is whether that research reduces the data or experimentation needed for a useful agent.

Credit and checking belong in the same story

According to OpenAI’s account of the mathematical work, OpenAI’s September account discusses Alpöge’s work and acknowledges priority concerning a forced-Euler result.

I would keep that interested account separate from an independent resolution of the mathematics or credit dispute. AI-assisted research needs an inspectable chain of attribution as well as a checked result.

The hardware can select the winning idea

According to Hooker’s Hardware Lottery essay, Sara Hooker examines how available hardware and software can favor particular research directions.

I find that framework useful when evaluating supposedly inevitable AI architectures. Technical success can reflect an ecosystem’s existing investments as well as the merits of the algorithm, which affects who can compete.

Research leadership meets enterprise constraints

According to Cohere’s executive-team page, Cohere lists Joelle Pineau as Chief AI Officer and describes an enterprise-focused AI platform.

The question for 2027 is whether research becomes useful inside organizations with private data and operational constraints. The decisive evidence would be dependable deployment and repeatable customer value, rather than a new model alone.

Follow the company: Who owns Cohere?

A business around an open inference engine

According to Inferact’s launch announcement, Inferact names Woosuk Kwon among its founders and announces $150M raised at an $800M valuation on January 22, 2026.

What matters next is how commercial engineering support interacts with the open project. Sponsoring an engine is different from owning all of its software, community or downstream value.

Make adaptation cheaper to attempt

According to Y Combinator’s Unsloth record, the company record names Daniel Han as a founder and describes open-source reinforcement learning and fine-tuning tools.

The question for 2027 is whether more teams can adapt models without owning a large training organization. Lower setup cost matters, but the trained model still needs a reliable evaluation and a deployment budget.

Program the system, then measure it

According to MIT’s faculty record, MIT lists Omar Khattab as an assistant professor in EECS.

His associated DSPy research direction puts attention on building and optimizing language-model programs. I am watching whether systematic evaluation can reduce the repeated prompt tweaking that makes agent projects costly to maintain.

Further records: DSPy official repository

The model definition is part of the infrastructure

According to The Transformers v5 announcement, Lysandre is among the announcement’s authors; the post describes interoperability across training, inference and deployment tools.

My focus is the maintenance work that lets a model reach users without every team rebuilding its implementation. Compatibility can be economically valuable even when it is less visible than a launch benchmark.

Compute has to fit the developer’s workflow

According to Bernhardsson’s own biography, Bernhardsson identifies himself as Modal’s founder and CEO and describes earlier open-source work including Luigi and Annoy.

The business question concerns the gap between renting hardware and getting useful work done on it. A good compute platform reduces deployment and orchestration effort, but its value must survive comparison with direct cloud infrastructure.

A revenue run rate is not a profit figure

According to Fireworks’ investor-hosted July announcement, the announcement identifies Lin Qiao and reports a $1.505B round plus an annualized revenue run rate exceeding $1B.

I would separate those measures before drawing a conclusion about the business. Serving revenue can grow alongside substantial compute costs; the useful question is what remains after delivering each workload.

Inference is a production engineering job

According to Baseten’s team and engineering page, Baseten identifies Tuhin Srivastava as CEO and co-founder and describes deployment engineering under real traffic.

The question for 2027 is whether the company can make reliability and latency predictable when models and customer workloads change. That is where infrastructure earns recurring revenue, rather than simply providing access to a GPU.

Creative models need a distribution layer

According to fal’s company history, fal says Burkay Gur and Gorkem Yurtseven founded the company in 2021 to address AI infrastructure challenges.

The question for 2027 is whether a platform for generative media can make model access dependable enough for professional products. Distribution matters, but providers must distinguish their own contribution from the underlying models they serve.

Local deployment changes the control question

According to Y Combinator’s Ollama record, the record names Jeffrey Morgan as a founder of a tool for running language models.

What matters next is how local model access changes privacy, portability and the cost of experimentation. Downloads and developer enthusiasm are useful signals, but neither establishes an enterprise business by itself.

Search economics can shape the agent

According to Eskildsen’s own biography, Eskildsen identifies turbopuffer as a search engine for object storage that he co-founded in 2023.

My focus is the cost of retaining and retrieving the context agents need. A cheaper search layer could change which products are feasible, but latency and correctness still matter when the data is large or frequently updated.

Evaluation becomes an operating expense

According to Goyal’s Series B announcement, Goyal describes founding Braintrust after building evaluation tools at Impira and Figma; the February announcement states an $80M round.

The question for 2027 is whether evaluation becomes a repeatable part of deployment rather than a launch checklist. The business case is strongest when teams can connect a test result to a decision about reliability and cost.

Preference is useful, but it is a particular measure

According to Angelopoulos’s research page, Angelopoulos identifies his Arena role and lists work on human-preference evaluation and conformal prediction.

What matters next is how evaluation platforms explain their populations, tasks and uncertainty. A preference ranking can guide a purchase, but it is not a universal measure of correctness or suitability for every workflow.

Governance is infrastructure too

According to MCP’s October maintainer record, the October 1, 2026 record lists David Soria Parra and Den Delimarsky as lead maintainers.

The business question concerns the rules that let tools connect to agents without every integration becoming a special case. The shared maintenance structure is more useful evidence than describing one person as the owner of an industry protocol.

Specialization needs a durable workload

According to Etched’s leadership page, Etched identifies Gavin Uberti as co-founder and CEO in its leadership section.

My focus is the trade-off behind specialized AI hardware: advantages can be large if the workload stays suitable, but chips take time and capital to develop. The full system, not an isolated operation, must justify the commitment.

The compiler belongs in the chip story

According to MatX’s team and technical description, MatX names Reiner Pope as co-founder and CEO and describes work spanning silicon, compilers, kernels and models.

The question for 2027 is whether coordinated hardware and software design can improve useful training or serving economics. A company can be technically differentiated without the evidence yet establishing a production advantage.

Software determines whether alternative hardware is useful

According to AMD’s 2026 speaker biography, AMD’s event materials identify Sharon Zhou as AI Chief of Staff.

My focus is the practical software work around hardware choice. An alternative accelerator becomes useful when developers can run and maintain their workloads without rebuilding every part of the stack.

The document pipeline can be the product

According to KPMG’s LlamaIndex investment announcement, the May 2025 record identifies Jerry Liu and explains managed document parsing, ingestion and retrieval services.

The business question concerns the unglamorous preparation of enterprise data. An agent cannot use a document reliably if tables, structure or permissions are lost before the model sees it.

Creating the app is only the beginning

According to KTH’s 2025 innovation-award account, KTH identifies Anton Osika and Fabian Hedin as Lovable’s founders and describes natural-language app creation.

I am watching what happens after the first working prototype. A business application must remain understandable, secure and maintainable as users request changes. That continuing work determines whether creation becomes durable value.

Voice models become an enterprise service

According to Staniszewski’s May company update, the update identifies Mati Staniszewski and reports that ElevenLabs surpassed $500M ARR in the first four months of 2026.

My focus is the shift from compelling generated voices to dependable communication workflows. ARR is a company-reported operating measure; the economics also depend on retention, inference costs and how customers use the service.

Human expertise stays in the training loop

According to Mercor’s mission and leadership statement, Mercor names Brendan Foody and Adarsh Hiremath as co-founders and co-CEOs and describes an expert network for training and evaluation.

The business question concerns the economics of human judgment as a model input. Labor-market participation, payment volume and company revenue are different measures; each needs its own definition before comparison.

A support agent must finish the customer’s problem

According to Decagon’s company and founders page, Decagon identifies Jesse Zhang as co-founder and CEO and describes agents across customer-service channels.

I am watching resolution quality rather than how many conversations a bot handles. A low-cost answer can become expensive if it creates escalations, repeats mistakes or damages the customer relationship.

Software autonomy needs an operating model

According to Factory’s September financing announcement, Grinberg and Reyes announce a $200M financing at a $5B valuation on September 15, 2026, following the earlier Series C.

The business question concerns the cost of reviewing and maintaining agent-produced software. The later financing is a distinct company event, not a replacement date for the earlier round or a personal-wealth estimate.

Move from software creation to delegated work

According to Emergent’s Wingman launch account, the April 2026 account names Mukund Jha and describes an agent operating through messaging interfaces.

The question for 2027 is whether familiar distribution channels make useful automation easier to adopt. Access to messages and tools creates operational responsibilities that a successful demo does not settle.

Creative output needs a production workflow

According to Runway’s 2026 summit record, the event materials identify Cristóbal Valenzuela as co-founder and co-CEO.

The business question concerns the transition from isolated generated clips to controllable creative production. Professionals need consistency, revision and usable rights, so the economically relevant output is a finished project rather than a striking sample.

Employee liquidity is not growth capital

According to Clay’s January tender announcement, the post by Kareem Amin describes employees being able to sell up to $55M of shares at a $5B valuation.

I would keep that employee-liquidity transaction separate from primary financing. For the product, I am watching whether automation improves useful commercial outcomes rather than simply increasing outbound activity.

Automation needs ownership and maintenance

According to n8n’s official company record, the press page identifies Jan Oberhauser as founder and CEO and dates the company’s founding to 2019.

What matters next is how teams connect models to established workflows. The value comes from dependable orchestration, permissions and exception handling; adding an agent to a flow does not remove those responsibilities.

Efficient visual generation still needs control

According to Stanford’s Rombach speaker biography, the 2024 record identifies Rombach as a Black Forest Labs co-founder and describes work on latent diffusion and visual models.

What matters next is how efficient generation combines with controllable revisions. A professional user needs the intended result on demand, so the number of failed attempts belongs in the cost calculation.

A creative platform has to earn repeat use

According to SaaStr’s Mashrabov speaker record, the conference record identifies Mashrabov as Higgsfield’s CEO and co-founder.

The question for 2027 is whether creators return for a complete workflow, not just to try a fashionable effect. The event’s financial language mixes run rate and ARR, so I have excluded it from the comparable operating table.

Being visible in AI answers is a business problem

According to Profound’s September announcement, the record identifies James Cadwallader and announces a $180M Series D at a $1.8B valuation.

What matters next is how marketers distinguish an observed mention from a customer actually choosing the product. A measurement tool should connect its indicators to useful business outcomes, while acknowledging that model answers change.

An agent’s memory can help or mislead it

According to The Reflexion paper, Noah Shinn is the paper’s first listed author; the framework uses feedback stored as text rather than updating model weights.

The question for 2027 is whether that approach becomes a dependable ingredient in agent products. The research artifact establishes a contribution, rather than a later company-financing outcome.

Research can become a production tool

According to UCL’s Synthesia case study, UCL attributes underpinning visual-synthesis research to Lourdes Agapito’s team and describes her co-founding role.

The business question concerns the practical conversion of research into a tool businesses can use repeatedly. The economic question is how much production effort is removed while preserving consent, identity control and a useful final result.

The user needs to direct the result

According to Pika’s founders biography, Pika identifies Demi Guo as CEO and co-founder in its team material.

What matters next is how consumer-friendly creation tools turn an idea into a controllable output. The useful measure includes revision effort and repeat use, rather than generation count alone.

The music business is part of the model business

According to Suno’s dated product and partnership posts, the September 2026 posts identify Mikey Shulman and describe model releases and music-industry partnerships.

What matters next is how creative control, permissions and distribution evolve together. A music model’s usefulness to customers depends on what they can do with its outputs, not only whether a generated song sounds persuasive.

Integration is part of the product cost

According to Wonderful’s company and leadership material, Wonderful identifies Bar Winkler and describes an enterprise platform supported by local deployment teams.

My focus is the economics of implementation. A model can be portable while the surrounding enterprise workflow remains highly specific; the labor needed to connect systems and manage exceptions affects margins and scalability.

Further records: Wonderful's publicly indexed team biography

Meeting context should survive the meeting

According to Granola’s product update record, Chris Pedregal appears as the author of releases including meeting chat, Windows support and shared team context.

The question for 2027 is whether a note-taking product becomes a dependable source of organizational memory. The value increases only if people can retrieve context accurately and manage what is recorded, retained and shared.

Clinical documentation has to be faithful

According to Abridge’s Lipton biography, Abridge identifies Lipton’s scientific and technology leadership and describes his work on robust machine learning.

What matters next is how documentation systems expose mistakes and support clinician review. Time savings are economically important, but I would not infer a clinical benefit from a product’s adoption or fundraising alone.

The workflow extends beyond the transcript

According to Ambience’s September 2025 appointment record, the company announces Nikhil Buduma’s move from chief scientist to CEO and describes documentation, coding and clinical workflows.

The question for 2027 is whether broader integration produces useful operational outcomes without weakening oversight. A company's deployment statement is evidence of its approach, not independent proof of improved patient outcomes.

A cited answer still needs source discipline

According to Sequoia’s Ziegler founder record, Sequoia associates Zachary Ziegler with OpenEvidence and links its clinical-information product.

The question for 2027 is whether information retrieval can remain useful while making the basis of an answer inspectable. Citations are an entry point for checking; they do not by themselves establish that the answer accurately represents the evidence.

Useful alerts must fit the care team

According to Bayesian Health’s founding account, the company identifies Suchi Saria as its founder and describes real-time clinical signals drawn from patient data.

The business question concerns the relationship between a predictive model and the people expected to act on it. An alert’s business and clinical value depend on workflow, false alarms and the intervention it enables.

A designed molecule must pass an experiment

According to Index Ventures’ Chai investment account, Index identifies Joshua Meier among Chai’s co-founders and describes a molecular-design business.

What matters next is how computational proposals become experimentally useful candidates. A partner announcement or prediction result does not establish a successful medicine; laboratory and development evidence must carry the next stage of the story.

General-purpose robotics needs repeatable transfer

According to Hausman’s research and founding account, Hausman identifies himself as a Physical Intelligence co-founder and lists work on robot learning and general-purpose skills.

What matters next is how much new engineering a robot needs for each deployment. Shared models could improve scalability, but the practical test includes intervention, setup and maintenance rather than a curated task video.

Learning from the world rather than a perfect dataset

According to Pathak’s CMU research description, Pathak describes agents learning from raw sensory data and research spanning vision, learning and robotics.

The question for 2027 is whether those methods reduce dependence on expensive curated demonstrations. A robot that learns more economically still has to do useful work safely and reliably in the intended setting.

A household robot is a service commitment

According to 1X’s NEO launch announcement, the October 2025 company release identifies Bernt Børnich and presents NEO as a home-robot product.

I am watching what a customer can actually rely on after delivery. Hardware reliability, human assistance, privacy and maintenance all belong in the cost of providing a useful household robot.

Embodied reasoning needs an action test

According to Parada’s Gemini Robotics announcement, the March 2025 post by Carolina Parada introduces vision-language-action and embodied-reasoning models.

What matters next is how language and visual understanding become dependable physical actions. The relevant evidence includes task success, adaptation and safety under changed conditions, rather than language fluency alone.

Legal work needs traceable context

According to Legora’s 2026 ILTACON material, Legora identifies Max Junestrand as CEO and co-founder and describes a platform for legal work.

The question for 2027 is whether legal teams can inspect and revise the steps between source material and final output. Time savings matter only alongside source fidelity, confidentiality and the professional review still required.

Distribution and trust are part of the product

According to Harvey’s Azure deployment announcement, the March 2024 record identifies Gabriel Pereyra and describes deployment and distribution through Azure.

What matters next is how a specialized AI product becomes purchasable and usable within a large organization. Distribution can accelerate adoption, but the product still has to justify its place in a professional workflow.

Follow the company: Who owns Harvey AI?

Research is only one stage of the financial workflow

According to Sivulka’s FlashDocs acquisition account, the June 2025 post identifies Sivulka and describes adding slide creation to retrieval and analysis workflows.

My focus is the handoff from evidence to a deliverable. A financial team needs a defensible chain from source documents to analysis to presentation, and errors can enter at every stage.

The banking workflow is the unit of value

According to Stengel’s April company update, Gabe Stengel authors Rogo’s April 29, 2026 announcement of a $160M Series D.

The question for 2027 is whether financial-workflow automation reduces total review and execution effort. A company’s financing is a dated capital event; it does not establish the accuracy of its analysis or the productivity of its users.

Grounding intelligence in structured facts

According to Helsing’s research-award account, the January 2026 release identifies Bordes and describes his co-authored research on structured knowledge embeddings.

What matters next is how research translates into reliable systems under demanding operational conditions. Academic recognition establishes a contribution; deployment claims need separate evidence about testing, constraints and oversight.

Count validated findings, not just generated exploits

According to XBOW’s account of its security-testing work, the June 2025 post says automated findings were reviewed by its security team before submission under platform rules.

For Oege de Moor’s story, that human-review step is economically important. I am watching how useful discovery, false positives and responsible reporting combine, rather than treating a leaderboard headline as fully unattended operation.

Further records: Sequoia's interview with Oege de Moor

Materials research needs a physical feedback loop

According to Periodic’s launch account, the company describes its founding team and a plan to combine AI with autonomous physical laboratories.

For Ekin Dogus Cubuk, I am watching whether the loop produces reproducible materials results. The company’s financing is recorded once alongside Fedus; a shared company event does not become two independent pieces of funding evidence.

A named industrial deployment is a concrete lead

According to Orbital’s BASF ECMS agreement announcement, the August 2026 post names Jonathan Godwin and describes licensing its CurieOS materials platform to BASF ECMS.

I would follow that relationship into evidence of actual industrial use. A named agreement is more specific than a general market claim, but it does not disclose realized revenue or independent proof of improved research productivity.

Start with the learning problem

According to Khan Academy’s discussion of DiCerbo’s learning approach, the organization identifies DiCerbo and describes her advice to begin technology decisions with the problem being solved.

The question for 2027 is whether AI tools improve learning outcomes rather than simply usage. Education requires a clear account of the teacher’s role, the student’s needs and what the evaluation actually measures.

Driving intelligence needs evidence on the road

According to Kendall’s research and company page, Kendall identifies his Wayve role and lists work on embodied AI, uncertainty and autonomous-driving models.

The business question concerns the progression from research to dependable deployment. The economics depend on the operating conditions, supervision and service required, which cannot be established by a capable-looking driving clip.

An evaluator needs more than a dramatic example

According to Apollo’s team and research description, Apollo identifies Marius Hobbhahn as CEO and co-founder and describes research on AI scheming.

The business question concerns the path from a concerning behavior in an experiment to a useful evaluation or safeguard. The business relevance depends on repeatability, realistic conditions and a clear account of what the test detects.

Design for a model that may not cooperate

According to Redwood’s leadership record, Redwood identifies Buck Shlegeris as Chief Executive Officer of the AI safety and security research nonprofit.

I am watching control methods that test whether oversight can remain useful even when a system behaves unexpectedly. The practical question is what they detect, what they miss and how much capability they sacrifice.

A narrow change can have a broader consequence

According to Truthful AI’s research and team page, the nonprofit identifies Owain Evans as team lead and lists research on deception, hidden reasoning and emergent misalignment.

The question for 2027 is whether training changes produce unintended behavior beyond the targeted task. That matters to buyers who adapt a model for a narrow use and assume the rest of its behavior remains unchanged.

Policy work changes commercial constraints

According to FAI’s Ball appointment account, FAI says Ball joins OpenAI as Head of Strategic Futures on July 6 while retaining a non-resident fellowship.

I am watching where policy arguments become concrete operating rules. His move also changes the institutional perspective behind those arguments, which readers should distinguish from independent academic or public-interest analysis.

A threshold is an initial filter, not the whole risk assessment

According to Heim and Koessler’s compute-threshold paper, the authors argue that training compute can identify models for scrutiny but should not determine mitigations by itself.

What matters next is how resource-based policy adapts when algorithms and inference change. A measurable input can be useful while remaining an imperfect proxy for capability or harm.

Implementation can matter more than the headline law

According to OECD’s Zenner biography, OECD describes Zenner’s digital-policy work for MEP Axel Voss and links his work on foundation-model regulation.

What matters next is how technical distinctions enter implementation and oversight. Companies need to know what actually applies to their systems; a historical biography’s negotiation language should not be mistaken for the current legal timetable.

Policy research needs a documented chain of reasoning

According to CSET’s current Toner biography, CSET identifies Toner as Executive Director.

I am watching evidence about AI competition and governance that can survive inspection beyond its headline. The value lies in transparent sources and distinctions, especially where technical and geopolitical claims interact.

The investment thesis should survive the market cycle

According to Guo’s account of launching Conviction, Guo describes an AI-focused venture firm built around companies using foundation models.

I am watching which investments produce durable businesses rather than brief distribution advantages. A fund’s commitments are managed capital, not the investor’s personal wealth or proof of returns.

Make predictions inspectable, then score them

According to Benaich’s State of AI report introduction, Benaich describes annual predictions and returning to assess them in the State of AI report.

I am watching that accountability as much as the predictions themselves. An investor’s research can be useful while reflecting an investment perspective; readers should be able to distinguish evidence, selection and the author’s bets.

Further records: Air Street's investment staff record

Product experience meets capital allocation

According to Kleiner Perkins’ Braswell biography, the firm identifies Braswell as a partner and describes earlier engineering and product work at Scale AI.

What matters next is how that experience shapes judgments about infrastructure and applications. A named investment or board seat establishes a relationship, but it does not establish the investor’s personal stake or a realized return.

The physical-world bet has a different delivery schedule

According to Lux’s Isford biography, Lux identifies Isford as a partner investing across software, computation and the physical world.

The question for 2027 is whether investment horizons match the work needed to deliver hardware, robotics or biology. Those businesses can have different validation and capital requirements from a software product, even when all use AI.

Open weights are not an open training recipe

According to Lambert’s departure and research account, Lambert describes his work on Tülu and open post-training and says he is developing a new project after leaving Ai2.

The business question concerns the information needed to reproduce how a model was improved. A released checkpoint is useful, but it does not reveal every ingredient in the training process. The departure account does not name the new organization, so I would not infer a current employer from it.

Interviews are evidence of what someone says

According to Patel’s podcast description, Patel identifies himself as the podcast’s host and describes a format built around researched interviews.

My focus is the questions that expose assumptions about cost, capability and strategy. A first-person interview can establish a speaker’s view; it does not independently verify every forecast or private-company number stated in it.

The people applying models need a shared vocabulary

According to AI Engineer’s history and leadership page, the organization identifies Shawn Wang and dates his AI Engineer essay and the first summit to 2023.

What matters next is how a professional community turns fast-moving tools into repeatable practice. Conferences and newsletters can accelerate learning, but attendance or audience size does not establish engineering quality.

Implementation exposes what a diagram leaves out

According to Raschka’s account of his research practice, Raschka describes reading papers and reference implementations, then building smaller versions to understand their details.

I am watching that bridge between research and working code. It can help readers judge architecture and evaluation choices for themselves, rather than rely on a summary that hides important assumptions.

Further records: Raschka's LLM implementation repository