A breakdown of the thirtieth Bay Area Frontier Research Club session at SignalFire: two talks, six questions from the audience, and one conversation worth more than both talks. What is a world model, where has the frontier shifted in 2026, where is the wall for both speakers, and why is this wall our job. Every statement is separated from opinion; anything unconfirmed is explicitly named.
The evening hinged on two questions, both posed by the organizers beforehand: what should a world model learn to predict, and what can an agent do when it has learned to imagine where its actions lead. The first question was taken from a report by OpenAI, the second from a report by Berkeley and Meta FAIR.
The analysis boils down to three levels, and it's useful to keep them separate because different people are responsible for them.
The first two levels are what was discussed on stage: prediction and choice. Both are probabilistic in nature, and both speakers hit the same wall: how to know if a step was correct, even if the accuracy is almost complete, but the remainder is unacceptable. The third level, verification, was not mentioned once in the room. This is precisely the layer on which ARLI stands: not "the model is confident," but "here's the record, here's the signature, check it yourself." The formulation worth taking away: the world provides the prediction, the record provides the verification.
Second conclusion, less obvious. The OpenAI speaker explicitly stated that general training does not provide the necessary accuracy in contact and friction physics, and that this is solved with domain-specific data and physical work. A live company in Shenzhen that seeks suppliers, handles acceptance and shipment, produces precisely this data. That is, our combination of “software and hands on the ground” ceases to be an explanation for an investor and becomes an answer to a technical problem named from their stage.
The notes are machine-generated and fragmented: model names, generation numbers, and some of the figures within them are off. Therefore, everything I present as fact is taken from articles and official communications, and everything heard in the hall is marked as spoken. Numbers that are not confirmed by a source are placed in a separate table and are not included in the materials. This is not a formality: two reports were about how expensive an unverified prediction is.
Event: the thirtieth session of the Bay Area Frontier Research Club, titled “Worlds for Agents: world models and the agents that learn inside them.” The format is honestly declared: dinner and article review, not reports for the sake of reports. Venue: SignalFire, November 4, 2026, from 5:30 PM to 8:30 PM. Organizer: the frontier syndicate, a nomadic program, partners are universities and funds.
| Parameter | As is |
|---|---|
| Session Title | Worlds for Agents: world models and the agents that learn inside them |
| Number | thirtieth session |
| Location | SignalFire, San Francisco |
| Date | November 4, 2026 |
| Format | two presentations of 10 or 15 minutes, followed by questions up to 20 minutes, reading materials distributed in advance |
| First presentation | Bonnie Li, OpenAI researcher, formerly Google DeepMind, participant in Genie 2 and SIMA 2 projects |
| Second presentation | Jiaxin Ge, Berkeley and Meta FAIR, lead author of the Agima paper and co-author of RATs, recognized at CoRL 2026 |
It is worth separately noting the organizers' request, because it is the door for us. The program calls itself grassroots and asks for three things: a venue for the event, a speaker with interesting work, and someone to pay for dinner. Plus, they maintain a calendar in Luma and publish announcements in advance. The person who announced this specifically said that there were more people wishing to attend than available spots, and that some people were left outside.
From the host's speech and the questions: SignalFire and Google Ventures as platforms and investors, Stanford and Berkeley as universities, Google DeepMind and Meta FAIR labs as employers of the speakers, plus practitioners: a person who builds evaluation infrastructure for sensitive areas, people from Gemini audio teams, and a team that launches agents for phone calls. This is not a conference; it's a close circle where half of those present work on the same floor as the speakers.
The definition given from the stage sounds like this: a world model is something that predicts what will happen next. It takes the state of the world and an action as input, and outputs the next state. The meaning for the agent is singular and practical: learning in imagination. The agent tries many things within the model before taking a step in the real world, and the real world is slow, expensive, and dangerous.
A world model is a prediction, and a prediction is probabilistic. This means any construct based on it inherits the probability of error, and in sensitive areas, this is not enough even with almost complete accuracy. This is precisely the question asked from the audience, and it remained without a definitive answer. Keep this in mind until the end of the document: it explains why our layer exists.
The speaker cited Dreamer as a complete example of how a world model is structured entirely, and this is convenient because all subsequent improvements are variations of these same three parts. The first goal: representation, which is a compressed state code that learns to reconstruct the original image from it. The second: transition, which is to predict the next state given the current state and action. The third: reward prediction. Learning occurs through image reconstruction, and this reconstruction signal is the main source of learning.
| Part | What it does | Why it's needed |
|---|---|---|
| Representation | compresses observation into a latent code via an encoder and decoder | to work not with pixels, but with a compact state |
| Transition | predicts the next state based on the current state and action | this is the world model itself; imaginary trajectories are rolled out through it |
| Reward | predicts how good the outcome is | so the agent chooses actions not randomly, but based on anticipated benefit |
Further, the entire construct works as follows: the agent represents the observation in a latent form, rolls out imaginary trajectories only through the transition, and the actor learns on these imaginary trajectories. Then the actor acts in the real world, the collected data is returned to the model, and the cycle repeats. The speaker named the original source and, from memory, pointed to a work from approximately six years ago. I checked: it's an article on behavior learned in latent imagination, December 2019, and that same line has a later continuation in the form of a generalized model for various control tasks.
Because each solves a separate problem: without representation, there is no compression; without transition, there is no imagination; without reward, there is no choice. All improvements in recent years are attacks on the first part, that is, on representation. There are two directions, and the next section is dedicated to them.
The question, which the speaker called the main direction of the year, is as follows: is it possible not to restore the image at all and learn directly in the latent space? The reason is simple: restoration forces the model to expend effort on everything visible, including what is irrelevant to the solution. In a complex image, extraneous elements begin to dominate the learning task and pull the representation away from what is needed for control.
There are two paths, and both were named from the stage.
| Path | How it works | Problem |
|---|---|---|
| Prediction in latent space without reconstruction | the first goal is the same: predict the next state from the current state and action | without the second goal, the model degenerates: it predicts a constant, and the error becomes zero. A regularizer is needed that prevents representations from collapsing into a single point |
| Contrastive learning | instead of reconstruction, mutual information between latent states is maximized: states close in time are brought closer, different ones are pushed apart | collapse does not occur by itself, because the task separates different examples; however, dependence on the quality of local features appears |
Regarding the regularizer and degeneration, the speaker explained honestly and step-by-step: if the only goal is to predict the next latent vector, then the cheapest solution is to always output the same vector, and the error will become zero, while the model learns nothing. Therefore, a second term is needed that separates the representations. And in the contrastive variant, this problem does not exist by design, because the task already requires distinguishing.
This is a case where the numbers drift in the recording, but everything is stationary in the source. The article is titled Contrastive World Models, one author is listed, who is the speaker herself, submitted on August 27, 2026. I am presenting the result based on the text of the article, not by ear.
| What has been done | How exactly |
|---|---|
| Foundation | Dreamer is taken, meaning a three-objective construct, and observation reconstruction is removed from it |
| What it was replaced with | a lower bound in the spirit of one that maximizes mutual information between state-action sequences and local fragments of future observation |
| What does this provide | the representation preserves what predicts the future and is not obligated to reconstruct what is irrelevant |
| Where tested | three modes with increasing visual complexity |
| Result | in the simple mode, it matches Dreamer and the momentum prediction baseline; with extraneous details and with a real video background, it significantly outperforms both |
| Side benefit | training is cheaper because the image decoder is entirely removed |
The last conclusion of the article is stated explicitly: robustness to visual clutter is especially important for transferring an agent from training to the real world. This is the answer to the almost rhetorical question from the audience about why world models poorly maintain contact and friction: part of the problem is not in physics, but in the representation spending effort on extraneous matters. By removing reconstruction, you return the model's attention to what predicts the future. The applied meaning for us is that this line directly leads to robots and manual work, not to beautiful pictures.
And one honest detail. The experiments are called small-scale, and the author herself does not hide this. That is, this is a result of a direction, not a finished product. For a conversation with an investor, this means: the topic is alive, but bets on it are made at the level of publications, not shipments.
The speaker went through the generations, and here we were lucky: dates and characteristics are confirmed by official communications, so the names that got distorted in the recording are accurately restored.
| Generation | Key, according to official reports |
|---|---|
| Genie 2 | large foundational world model: generates diverse interactive environments, action control, all dynamics learned in a latent space |
| Genie 3 | general purpose: generates dynamic worlds from a text prompt, in which one can walk in real-time, 24 frames per second, several minutes of consistency, 720 pixels resolution |
| SIMA 1 | was able to perform over six hundred skills according to instructions in various games, acting like a human: looks at the screen, operates keyboard and mouse, without access to the internal game mechanics |
| SIMA 2 | Gemini became the internal core: the agent not only executes instructions but reasons about goals, converses, and improves over time |
What the speaker showed as her favorite consistency demonstration was painting a wall: the agent paints a section, moves to another location, returns, and everything remains exactly as it was left. She specifically emphasized that the model does not have an explicit map, and memory of the room emerges from large-scale training. This is the strongest part of the story: not the image, but the fact that the model holds the world in its head without explicit structure.
The organizers formulated this in the announcement with one phrase, and it deserves a separate mention: the agent continued to improve within model-generated worlds, without new human demonstrations. That is, the loop closed: the world is produced by the model, the agent acts within it, the attempt is evaluated, and the cycle continues on its own.
The first consequence is good and is sold from the stage: training becomes cheaper because neither labeling, nor people, nor real robots are needed at each step. The second consequence is inconvenient: if the human is removed from the loop, then the one who confirms that the result is correct also disappears. The evaluation is set by the world itself, meaning the same probabilistic construct as the agent. This is precisely the wall that was discussed in the questions: how to know that the step was correct if it's checked by a prediction, not a recording.
This is why the entire evening revolved around reliability, although not a single slide was about reliability. Questions were asked about high stakes in medicine, extraneous details, contact physics, and the model inventing scenarios instead of failures. All these questions are one question: what confirms the result if only a prediction can be trusted. We have an answer to this question, and it differs in nature, not quality: a recording of the step, which can be checked by an outsider, does not predict, but fixes.
The first work of the second report is called Train An Imaginative Agent, project abbreviated Agima, submitted on October 19, 2026. The motivation in the article is formulated as follows: an outwardly intelligent push of a box can make a puzzle unsolvable, so the agent must anticipate consequences before execution, especially when feedback is rare and actions are irreversible.
| What the agent does | How it works |
|---|---|
| Proposes actions | one general model, working with both text and images, acts as the policy |
| Imagines the result | the same model generates an image of what will happen, i.e., acts as a world model |
| Evaluates options | decides what to execute by reasoning about the imagined future |
| Does not stick to one plan | can scroll through a sequence of steps, consider alternatives, and compare possible futures |
| How it learns | first supervised fine-tuning, then reinforcement learning |
Scenes on which this was tested: maze navigation to a viewpoint, pushing boxes, a pursuit game where you can't eat too much, and continuous control. Training was done on simple variants, and testing on complex ones, meaning the ability to transfer, not memorize, was checked. It is stated that the result on these tasks is comparable to or better than known closed models, and noticeably better than the same approach without imagination.
This part costs more than all the slides. After reinforcement learning, the agent spontaneously developed behavior that no one programmed: before training, the second plan was generated almost randomly and frequently, but after about a hundred iterations, it is only used when the first plan fails. At the same time, according to her, a mixed reasoning format appeared, which was simply not found in the data before training.
Computational savings emerged as a consequence of feedback, not as a rule written by a human. This is exactly what defends any of our discussions about the cost of a step: correct behavior arises where a signal is received for a bad choice. To translate into our language: if a step has a receipt showing that it failed, then the agent's accuracy grows precisely from this. A record is not reporting, it is a training signal.
I heard all the numbers from this part, but they are not confirmed in the text of the article available to me: this refers to the frequency of the second plan before and after training and the number of iterations. Therefore, until cross-referenced with the full version of the article, these numbers will not be included in the materials. In the document, they are left as stated from the stage, and this is noted in a separate table below.
The work is called Playful Agentic Robot Learning, submitted on June seventeenth, 2026, and noted at CoRL 2026 as a spotlight. The motivation is human: people acquire skills by playing, and the question is whether a robot can be given the same. The system is multi-agent, and this is perhaps the most interesting for us, because it is already about the division of labor between agents, and not about a single model.
| Role | What it does |
|---|---|
| Proposer | invents a task for the current environment, based on what it already knows |
| Selector | evaluates tasks by novelty and complexity: too simple is useless, too difficult is unsolvable |
| Planner | breaks down the task into steps and writes code that will be executed |
| Executor | runs the code in a real environment or in a simulator |
| Dissector | explains failures and corrects the approach |
| Memory | successful techniques are compiled into a skill library and remembered as behavior |
According to the speaker: an increase of about twenty percent from such playback, transfer of skills between different hands, and work on a real robot with growth from approximately thirty to thirty-five percent. I cite the numbers as heard, because they are not confirmed in the article text available to me. Separately, she said that they present their work on robots separately, at another venue, and this is another reason to keep it on the reading list.
And one observation that is easy to miss. She explicitly states that transferring this to continuous control is difficult, but to text and images it is easy, and that the question of increasing uncertainty over a long horizon remains open. That is, in the second report, the ceiling is the same: the further in time and the coarser the physics, the less confidence, and no amount of training cancels this out yet.
The evening's format is structured such that questions are the content itself: the talks are short, and the discussion is longer. Below are six topics that were raised, with answers as they were given.
| Question | Answer | What this means for us |
|---|---|---|
| Can an agent and a world model be trained together | yes, previously they were always trained together; on very large versions, no longer | co-training is a sign of small scale; on large scales, everything diverges |
| What is the training objective for the contrastive approach | maximization of mutual information, lower bound, closest work from 2018; prediction horizon can be changed, with zero horizon, comparison is between global and local features | this is a mature line, not a new fad, it can be referenced |
| Why models struggle with contact, friction, and manual tasks | domain-specific data is needed for a particular task on top of general training | direct confirmation of the role of our Chinese side, see the data section |
| How to make a model predict actual failures, rather than invent scenarios, in medicine | caution is justified; confidence comes from thorough training and a large number of checks, there is no guarantee | no guarantee is precisely the market gap for our layer |
| If accuracy is 99.99 percent, the remainder is still unacceptable, how to live with it | co-train policy and world model, collect data in the real world, continuously improve, and the limit remains | the strongest argument of the evening in favor of verification, not prediction |
| Will the agent not hack the reward by exploiting numerical instability in physics | this is a big problem everywhere; here we haven't seen it yet, because these works are early and the model is not overtrained | early stage question: as soon as the agent lives in a real environment, the question will return |
There was no answer regarding long-horizon uncertainty: it was admitted that it hadn't been checked. The answer regarding ultimate precision was honest, but essentially it was "we don't know how to guarantee it." The speaker refused to discuss model sizes and referred to open works by Chinese laboratories. Three questions, three gaps, and all three lie not in training, but in verification. This is precisely our area.
A person who builds agents working in sales on behalf of a company spoke from the audience. Here's what he said, and I'm quoting it almost verbatim, because every word here is impactful.
We have several digital employees working like humans. The salespeople on the other side didn't know they were talking to agents. They called and asked to be connected to a human. The problem we can't solve is that agents don't understand how to conduct a conversation when there are multiple parties in a deal: buyers, lawyers, accountants. A human understands the need to switch between them. We don't see how agents do this.
The speaker's response was friendly and general: there is work on multi-agent communication, and it will get better with new models, but the environment must be as close to reality as possible and properly configured.
The practitioner described a failure on two levels at once, and no one noticed the second level. First level: the agent doesn't know how to talk to multiple parties. Second, much more important: the salesperson on the other side didn't know they were talking to a machine, and no one recorded it. That is, work was done, money moved, but there was no trace by which this could be established. This is precisely the case for which our product with receipts exists, and it was presented in someone else's words, without a single slide of ours.
Hence the practical conclusion for our materials. We are not selling "a smarter agent" or "a more accurate model." We are selling something that a human, agent, or regulator can verify: who performed a step, when, under what key, and with what result. The remark from the audience is a ready example of the value of such a layer: if that salesperson had a record, the conversation about agent quality would have gone completely differently.
If you condense the entire evening, three walls will remain, and all three are about verification, not about training.
The first wall is inherently insurmountable: prediction will remain probabilistic, no matter how much it's improved. The second and third require what laboratories lack: real work on the ground and real accounting for parties. This means the demand for the verification layer will not disappear with the new generation of models; on the contrary, it will grow with them: the more work is delegated to agents, the more important it is to be able to show exactly what they did.
The evening was in a language where half the words don't mean what they seem. I compiled a short dictionary so that we don't get confused and don't confuse our interlocutor in our future materials.
| Word | What it actually means | What it is not |
|---|---|---|
| World Model | prediction of the next state based on state and action | not a simulator with physics and not a game engine |
| Training in imagination | the agent trains on its own predictions within the model | not training on real data and not self-training without a goal |
| Latent space | a compact numerical description of the state instead of an image | not "understanding" and not symbols |
| Collapse | degeneration: the model outputs the same thing and doesn't learn | not overfitting and not overtraining |
| Contrastive learning | bringing similar things closer and separating different things instead of reconstruction | not comparison with a human and not a benchmark |
| Imaginary agent | an agent that first simulates consequences, then acts | not a "dreamer": this is a waste of computation to avoid irreversible errors |
| Skill library | a repository of successful techniques gathered from the game | not retraining for each task anew |
| Verifiability | the ability for an outsider to verify the record | was not mentioned at all in this conversation |
The first conclusion is the positioning. All three pillars of the evening—probability, physics, and many parties—converge on one question: how is the result confirmed? The world model answers with a prediction, and a prediction is probabilistic by design. Our answer responds with a record: step, key, signature, third-party verification. This is not “better”; it's a different level of task, and that's precisely why it doesn't compete with their work but complements it.
The second conclusion is practical. The speaker from OpenAI stated directly: general training does not provide the necessary accuracy in contact and friction; domain data and work in this area are needed. This is what a living company in Shenzhen produces daily: verified suppliers, acceptance, shipment, deviations, corrections. This results in a rare thing for an investor conversation: our business and our answer to a technical problem are one and the same. We don't invent applications for data; they already exist as a byproduct of our work.
Not a single report mentioned a word about the market for agent services, inter-company settlements, receipts, trust, and agent evaluation. The room is occupied with training, environment, and physics. This means our topic is not native there, and the pitch must go through their language: reliability and verification as a condition without which their agent cannot be released beyond the laboratory. If we go straight into talking about our products, we will be listened to politely and forgotten by dinner.
It's useful to state this plainly, because the temptation to call our receipts a “world model for trust” is great, and it is also harmful. A receipt predicts nothing. It records: who performed the step, when, under what key, what bytes resulted, and allows an outsider to verify the signature. This is a different type of assertion about the world.
| Property | World Model | Step receipt |
|---|---|---|
| What it asserts | what will happen next | what happened and what it's related to |
| Nature of the assertion | probabilistic | deterministic |
| How it is verified | accuracy on tests | comparison of bytes and signature |
| By whom it is verified | the same team | any outsider |
| What it provides in case of error | a signal for model improvement | a trace by which the guilty step is identified |
| Limit | there always remains a probability | A signature proves bytes, not the truthfulness of a string. |
A signed receipt proves signatures and bytes, not the truthfulness of what is written in it. If the executor lied in the string and signed the lie, the receipt will not catch it; it will only show who signed it. We state this limitation first and before our interlocutor does, because the entire evening was about the high cost of undifferentiated certainty.
The question about contact and friction was asked by someone who, judging by the wording, works with robots manually. The answer was mundane and the most valuable: domain-specific data and fine-tuning for a specific task are needed; general training does not achieve the required accuracy. This is not the speaker's opinion; it is a recognized deficit in the field.
Now let's look at what the Chinese company produces. It searches for suppliers and screens out bad ones. It conducts acceptance and records deviations. It organizes shipments and catches missed deadlines. It collects price lists, conducts correspondence, and processes documents. Each of these steps is a labeled example of what manual work looks like, where mistakes were made, and how they were corrected. This is precisely the type of data that laboratories lack and that cannot be collected from open sources.
Physical intelligence needs a machine-readable production chain. Robots are becoming autonomous, but the chains behind them are not. We produce this chain in two forms simultaneously: as a verifiable step record and as labeled data of real physical work. The second is a byproduct of the first, and that is why it is cheaper for us than for someone who builds such data from scratch.
And a caveat immediately, so that no one catches us exaggerating. We have little data today: it is data from one small company, not a stream. Therefore, the honest formulation is not “we have a dataset,” but “we have a way to obtain such data with the same work for which clients are already paying.”
The room speaks the language of learning: representation, prediction, domain, episode, reward. If we come with words about the market, receipts, and contracts, we will be politely listened to and forgotten. Therefore, the entry is built from their language, and we now have a ready bridge for this: their own three walls.
| Their language | Our translation | What we present |
|---|---|---|
| Evaluation of an attempt in the environment | a record that an attempt actually occurred and what its outcome was | a signed receipt of the step and the possibility of verification by a third party |
| Domain data for contact and friction | labeled examples of real physical work | work of a Chinese company: suppliers, acceptance, rejections |
| Multi-agent communication | it is visible which participants did what and for what they are responsible | step log with authors and keys |
| Reliability at high stakes | verification instead of trusting a prediction | fingerprints, signatures, analysis in case of substitution |
We don't improve prediction; we make work verifiable. If you need an agent to learn better, that's for tonight's speakers. If you need an outsider to be able to establish a year from now what the agent did and on whose behalf, that's for us. The second is not replaced by the first, no matter how many models you train.
Section that makes this document usable. Below, all essential information is divided into three groups: confirmed by an official source, confirmed by the article text, heard from the stage and unconfirmed.
| Assertion | Status | What it's based on |
|---|---|---|
| World model predicts the next state based on state and action | confirmed | wording in the article text and session announcement |
| Contrastive approach outperforms reconstruction with extraneous details and video background | confirmed | article text on contrastive world models, submitted August 27, 2026 |
| Genie 3 operates in real-time, 24 frames per second, consistency for several minutes | confirmed | official announcement about Genie 3 from August 5, 2025 |
| SIMA 2 integrated Gemini and can reason about goals and improve over time | confirmed | official announcement about SIMA 2 from November 13, 2025 |
| Agent improved within generated worlds without new demonstrations | confirmed | session announcement by organizers |
| Dreamer as an example with three learning objectives | confirmed | article on behavior learned in latent imagination, December 2019 |
| The second plan is adopted only if the first fails, after reinforcement learning | heard | from the stage; numbers and frequency not confirmed in the available article text |
| An increase of about twenty percent from the robot playing | heard | from the stage; not confirmed in the article text |
| Growth from thirty to thirty-five percent on a real robot | heard | from the stage; requires verification with the full version of the article |
| Agents on calls, salespeople didn't know they were talking to a machine | heard | remark from the audience, third-party assertion without a source |
| Model sizes | not disclosed | the speaker refused to answer and referred to open works by Chinese laboratories |
Contrastive World Models, sole author is the speaker, August 2026. Genie 2 as a large foundational world model, December 2024. SIMA 2 as a general embodied agent for virtual worlds, November 2025. Training an Imagining Agent, October 2026. Game Agent Learning for Robots, June 2026, noted at CoRL. All five of these constitute the evening's curriculum, and they should be kept in one folder.
A separate section on how I handled the transcription, because this is a check on the honesty of the entire document. The recording is machine-generated: it swallows endings, confuses model names, and combines numerals into random combinations. An example from parsing practice: in one place, instead of a model generation name, a non-existent word was uttered, and in another, percentage increases merged with an adjacent phrase.
| What has been done | Why |
|---|---|
| Generation names restored from official communications, not from the recording | in the recording, they are distorted beyond recognition, and an error here would cost the entire document its credibility |
| Characteristics are taken from primary sources with dates | so that any number can be re-verified in a minute |
| Numbers from the second report are marked as heard | they are not confirmed in the available text of the article |
| Direct quotes are given by meaning, without embellishment | the transcript is fragmented, a verbatim rendition would be a fake of accuracy |
Who exactly was in the hall, besides those named by the host. The name of the company from which the person with agents on calls came. The exact composition of participants and who funds the program. All of this is in the list of tasks, not in the conclusions, because speculation here is useless and harmful.
The organizers listed directly from the stage what they need, and this is a rarity: usually such things are found out through correspondence. They need a venue for the event, a speaker with interesting work, and someone to pay for dinner. Plus, the calendar is kept open and announcements are published in advance, and there are fewer spots than interested parties.
| Door | What gives us | Entry price |
|---|---|---|
| Host a session at our venue | a spot in their schedule and an introduction to the entire program | evening organization and dinner |
| Propose a talk | a conversation about verifiability in their language, with their audience | finished material for twenty minutes and an article |
| Pay for the session dinner | mention and an occasion for acquaintance without a talk | cost of dinner, the lowest of three |
| Come again as a participant | acquaintances and updating the picture once every one and a half weeks | ticket and time; but spots are selling out |
And what requires speed: the next robotics session is already at a different venue, and it was mentioned in the recording that work on robots is presented separately. If we want to join this circle not as spectators, the right move is not a letter telling about ourselves, but an offer of specific benefit: a venue plus a talk about how an agent's work is verified, with their vocabulary and on their tasks.
The plan is short because the evening provided not tasks, but language. Three steps, each done in one go.
| Step | What exactly | To whom and why |
|---|---|---|
| One-page note | prediction is probabilistic, record is deterministic, physics cannot be learned without domain data, which is provided by hands-on work | the basis of writing and speaking; the support consists of three questions from the audience and a conclusion about domain data |
| Reconciliation of two articles | read the entire imagining agent learning and robot game learning and write down the numbers | to ourselves, so that no number in the materials is merely heard |
| Letter to the program | propose a venue and a session on agent work verifiability | to the organizers, with their own words about the venue and the speaker |
Check the numbers against the two articles. Understand what company brought the agents to the calls: it's either a client or a competitor, and in both cases, it's more useful than any new conversation. Get the list of participants and the calendar. And determine who can host the evening for us: the venue is a physical asset, and it's not yet resolved for us in San Francisco.
This analysis is based on the transcript and open sources. Below is what it does not replace.
| Gap | Why it's important |
|---|---|
| Full texts of two articles not read entirely | the numbers of the second report are currently heard, and until reconciliation, they are not included in the materials |
| It has not been established who exactly was in the hall | networking is half the value of such evenings, and the list must be obtained, not guessed |
| No full recording of questions | some remarks in the transcript are cut off, and I draw no conclusions from them |
| The position of the verification laboratories has not been clarified | no one asked directly how they themselves confirm the agent's result; this is the best question for the next session |
It is not clear from the transcript whether the question about accounting for completed steps was even raised in the hall. It is not clear whether anything substantial was discussed at dinner, which is usually more important than the reports. It is not clear who reacted and how to the remark about agents on calls. All of this requires either your memory or a subsequent visit, and I do not substitute guesses for facts.
Should we attend such evenings? Yes, but not as spectators. The value is not in the reports, but in three things: the remark about agents on calls, the recognition of the deficit of domain data, and the organizers' open request for a venue and a speaker.
Which of these is closest to our products. Attempt evaluation in the environment: they use a model for it, we record it. The same word, two different mechanisms.
What of this is most dangerous for us. Nothing. Their work strengthens our arguments, because no one has stated more honestly than them that prediction offers no guarantee.
What to expect next. The next session on robots, a separate presentation of the work on a playing robot, and likely a continuation of the line on contrastive world models. If we get on the calendar now, by tomorrow evening we will have not a story about ourselves, but our own topic in their language.
They taught the agent to imagine consequences. We make it so that consequences can be verified. The latter is not replaced by the former and does not depend on which model comes out next.
One site of yours, measured, with the price for the work in the same document
ARLI, Inc., a Delaware corporation. Questions: hq@arliwork.com.