How Close Are We to B-Bots? Ron's Gone Wrong, Five Years On

In 2021 an animated film imagined a world where every child gets a robot best friend in the post. It talks, it follows you to school, it knows what you like, and it finds you friends who like the same things. One lonely boy gets a broken one, and the broken one turns out to be the only real friend in the film. Five years later, the robot’s mind is already in our pockets, and on 21 October 2026 the head of the world’s most valuable chip company stood on a stage in Berlin playing with its distant cousins: four little walking ducks. So how close are we to B-Bots?

Ron’s Gone Wrong came from Locksmith Animation, the studio’s first feature, directed by Sarah Smith and Jean-Philippe Vine and released by 20th Century Studios on 22 October 2021, a year before ChatGPT reached the public. Smith has said she wanted to make a film about the internet’s effect on children’s friendships and self-esteem. It is not really a film about robots. It is a film about what happens when a social-media company decides it can sell friendship.

My argument is this: the B-Bot’s mind is largely here, already running in phones, and its body is arriving as small, cheap robots trained in simulation. The hard gap is the one the film was about: a companion built to outlive its maker, keep a child’s data safe, and hand the child back to other people. That is a design and business-model problem more than a hardware one, and we should solve it before the hardware catches up.

What follows recaps the film (spoiler-light until a marked section), looks at the obvious real-world comparison, breaks a B-Bot into the parts we would need to build, and checks each part against the 2026 evidence: phones, the Microducks, home robots, the research on AI companionship and the new laws. Then it asks which kinds of AI would close the gap, and what kind of friend we should want at the end of it.

Play · YouTube

The film: a best friend, out of the box

This section covers the setup only.

In the film, the tech giant Bubble launches the B-Bot, a walking, talking robot companion for children. The pitch, in the words of the film’s own mock commercial, is a “best friend, out of the box”. Each B-Bot connects to its child’s likes, dislikes and social media posts, builds a profile, and uses a friendship algorithm to match the child with other children who share their interests. Soon almost every child at school has one.

Barney Pudowski, a socially awkward 12-year-old, is the exception, until his family gets hold of a B-Bot that fell from a delivery van. It is damaged. It has not had all of its settings installed, its safety features are missing, and it does not know how to be a friend. Barney names it Ron after its model number. Ron’s attempts to learn friendship from scratch drive the comedy.

Two executives sit behind the product. Marc Wydell, the young founder and chief executive (Justice Smith), whom /Film calls the company’s “wunderkind CEO”, sincerely believes B-Bots will help lonely children. Andrew Morris (Rob Delaney), a profit-minded executive whom the same review describes as “designed to look somewhat like Steve Jobs or Tim Cook”, sees Ron as a public-relations disaster and wants him destroyed. The film’s darkest joke is that the robots also gather data: Andrew is eventually caught admitting that they spy on their owners for profit.

Marc, or Mark?

The film never names a real person. But a social-technology company, a young founder named Marc and a product sold as a cure for children’s loneliness make the comparison hard to avoid, and real life has since made it closer.

On 29 April 2025, Mark Zuckerberg told the Dwarkesh Podcast that “the average American has fewer than three friends”, that “the average person has demand for meaningfully more”, and that the number was “something like 15 friends”. He said he expected people to “use AI for a lot of these social tasks”, while adding that his default view is that AI will probably not replace in-person connection. Critics pointed out the irony of a social-media founder offering AI as the fix for a loneliness problem social media is often blamed for. Read next to the film, it is almost the B-Bot pitch: a company that already knows who your friends are offering to supply the ones you are missing.

The body is on Meta’s agenda too. In February 2025 Reuters reported an internal memo from Meta’s chief technology officer, Andrew Bosworth, setting up a robotics group inside Reality Labs focused on “consumer humanoid robots” and on making Meta’s Llama models the platform for them. The group’s leader, hired from the self-driving company Cruise, is called Marc Whitten. Fiction sometimes rhymes.

Context: why this question is live now

For most of the five years since the film, the honest answer to “how close are we?” was “the talking part is close, the robot part is not”. Two things changed that in 2025 and 2026: language models small enough to run on a phone, and robot bodies cheap enough to sell as Christmas presents.

Exhibit 1

20K+Microduck preorders within a day of the GTC Berlin keynoteTECHCRUNCH · OCT 2026

$399introductory price of the 25 cm walking MicroduckTHE NEXT WEB · AUG 2026

1 GBRAM on board the Microduck: enough to walk, not to hold a conversationTHE NEXT WEB · AUG 2026

~3Bparameters in Apple’s on-device language model, which runs offline inside appsAPPLE · JUL 2025

Preorder figure from Hugging Face's chief executive on X, as reported by TechCrunch; it is a company claim. Microduck specifications are provisional until it ships.

Source: TechCrunch (23 Oct 2026); The Next Web (Aug 2026); Apple Foundation Models tech report (Jul 2025)

The timeline below shows how quickly the pieces have moved, and how quickly the warnings have arrived alongside them.

Exhibit 2

1. OCT 2021Ron’s Gone Wrong opensA social-tech giant sells every child a robot best friend.
2. DEC 2024 – JAN 2025Moxie goes darkEmbodied shuts down; its $799 child companion robot stops working when the cloud servers switch off.
3. FEB 2025Meta forms a humanoid robotics groupInside Reality Labs, aimed at consumer humanoids running Meta’s Llama models.
4. MAR 2025MIT and OpenAI study AI companionshipHeavier chatbot users show worse loneliness and socialising outcomes.
5. APR 2025“Fewer than three friends”Zuckerberg says people want about 15 and AI can help.
6. JUL 2025Reachy Mini, a desktop robot for AI buildersHugging Face’s first robot; more than 10,000 sold within about a year.
7. OCT 2025Character.AI ends open-ended chat for under-18sAfter lawsuits over harm to teenagers.
8. JAN 2026California’s companion-chatbot law takes effectSB 243: tell minors they are talking to an AI, remind them to take breaks.
9. AUG – SEP 2026Microduck preorders; NVIDIA agrees to buy Hugging FaceA $399 walking robot, then a $12.9 billion deal for the company behind it. California signs a toy moratorium.
10. 21 OCT 2026Jensen Huang plays with Microducks at GTC BerlinPreorders pass 20,000 the next day.
11. 1 JAN 2027California’s companion-toy moratorium beginsNo toys with companion chatbots for under-16s until 2031.

Source: Wikipedia; The Register; Reuters via CTV; MIT Media Lab; Dwarkesh Podcast; Hugging Face; Bloomberg; CalMatters; The Next Web; NVIDIA SEC filing; TechCrunch

A framework: mind, body and bond

To answer “how close are we?” properly, it helps to break a B-Bot into the parts someone would have to build. I use three layers, each with a few capabilities.

• Mind is what the robot says and knows: holding a conversation, remembering the child, recognising faces and places.
• Body is what it does in the physical world: moving around with a child all day, handling objects safely, and running long enough on a battery without its maker’s servers.
• Bond is what it is for: keeping a child safe, protecting their data, and being a good friend, which, as the film argues, includes helping the child make human friends.

Exhibit 3

	In the film	Closest thing in late 2026	Verdict
Mind: talk	Chats, jokes and keeps its child company	Voice assistants; on-device models in phones that run offline	Here
Mind: know you	Builds a profile from likes, dislikes and posts	Assistant memory features; phones already hold the data	Here, which is the worry
Body: move	Goes everywhere with its child, including school	Microduck walks and gets up after falls, for about an hour per battery	Arriving, small and short-lived
Body: handle the world	Plays, carries, steers a child out of trouble	Microduck lifts about 100 g; 1X NEO leans on remote human operators	Not yet
Body: keep running	Updated from Bubble’s cloud, never fails (until Ron)	Moxie stopped working when its maker shut its servers	Not yet
Bond: keep a child safe	Safety settings, which Ron is missing	Disclosure and break rules in California; Character.AI’s under-18 ban	Contested
Bond: be a friend	A friendship algorithm matches children by interests	Heavy chatbot use goes with worse loneliness in a 2025 trial	Unsolved

Framework by the author. Evidence is drawn from the sources cited in this article. The table scrolls sideways on small screens.

If I turn those verdicts into a rough score, the answer to the title looks like this.

Exhibit 4

Mind: talk~85%

Mind: know you~70%

Body: move~35%

Body: keep running~20%

Body: handle the world~15%

Bond: be a good friend~10%

Author's judgement, not a measurement. It reflects products a family can buy or preorder in late 2026, not lab demonstrations.

View data table

Capability	Estimate
Talk	about 85%
Know you	about 70%
Move	about 35%
Keep running	about 20%
Handle the world	about 15%
Be a good friend	about 10%

The rest of the article tests those numbers, layer by layer.

Deep dive 1: the mind is already in your pocket

A B-Bot’s conversation, memory and recognition are the parts that have arrived fastest, and they arrived in phones rather than robots.

Apple’s 2025 technical report describes an on-device language model of about 3 billion parameters, compressed to run on recent iPhones, iPads and Macs. Through Apple’s Foundation Models framework, app developers can call it directly: it works offline, keeps data on the device and costs the developer nothing per request. Google offers the equivalent on Android through Gemini Nano, which runs on supported phones through the system’s AICore service. Both companies send harder requests to larger models in the cloud, but everyday conversation, summarising, and pulling facts out of messages now happen on the device.

That matters for a B-Bot in three ways:

1. Latency. A companion that pauses for two seconds before answering does not feel like a friend. Local models answer immediately and work without signal, on a bus or in a basement.
2. Privacy. The film’s B-Bot sends everything to Bubble. A local model can know a child well without the company knowing them at all. That is the difference between a diary and a data pipeline.
3. Survival. A mind that lives on the device does not die when the company does. Remember that for the Moxie story below.

The phone also already carries most of what a B-Bot would need to know you: your photos, your messages, your calendar, your music and who you talk to most. The film imagined Bubble collecting that data. In reality it is already collected, in one device, by the companies that make phones. A B-Bot does not need a new data pipeline. It needs permission to use the existing one, which is why the “know you” score is high and why it is the part to worry about.

What this means for the framework: the mind is close enough that the bottleneck has moved. The question is no longer “can it talk?” but “who does it talk to about you?”

Deep dive 2: the body is arriving, small first

On 21 October 2026, during his keynote at NVIDIA’s GTC Berlin conference, Jensen Huang spent part of his time on stage with four small duck-shaped robots. In the video that went viral afterwards, they stomp their feet, make R2-D2-style noises, sit, fall over for a “nap” and do somersaults on voice command. The next day Clément Delangue, the chief executive of Hugging Face, said preorders had passed 20,000.

The Microduck is made by Pollen Robotics, the French company Hugging Face bought in 2025, and it is the company’s second robot after Reachy Mini, a desktop robot launched in July 2025 that had sold more than 10,000 units by August 2026. In September 2026 NVIDIA agreed to buy Hugging Face for about $12.9 billion, according to its SEC filing. The company whose chips train most of the world’s AI models is buying the company that sells one of the cheapest walking robots you can preorder.

Exhibit 5

	Microduck	What it means for a B-Bot
Size and price	25 cm, under 800 g, $399 introductory	A robot body now costs less than a flagship phone
Moving	15 motors; walks, crouches, gets back up after falling; optional roller-skate feet	Getting around and recovering from a fall work at toy scale
Sensing	One camera, a small time-of-flight depth sensor, two motion sensors	Enough to see and avoid; not enough to understand a room
Handling	Picks up small objects with its beak; about 100 g	Hands are the missing piece
Computing	Rockchip RK3566 with 1 GB of RAM; a 50 Hz control loop	The body thinks fast but small; conversation must live elsewhere
Battery	Removable camera battery, about one hour	A B-Bot needs a school day, not a lunch break
Learning	Skills learned in simulation with reinforcement learning, then moved to the real robot; owners can teach and share new tricks	This is how robot bodies will improve fastest

Some specifications are provisional until shipping. The software is open source; Pollen says the hardware design is not. The table scrolls sideways on small screens.

Source: The Next Web (Aug 2026); Axios (27 Aug 2026); TechCrunch (23 Oct 2026)

Three things stand out.

The brain and the body are separate. With 1 GB of RAM, the Microduck can run the small control model that keeps it upright, 50 times a second, but not a language model you could talk to. It makes sounds instead: Pollen describes them as “weird little sounds, closer to a creature than an assistant”. Any conversation will have to come from a phone or the cloud. That is the architecture of the near future: a body as a peripheral of the phone, the way earbuds are.

The skills came from simulation. The ducks shipped knowing how to walk, sit and stand because they practised in a simulated world first, using NVIDIA’s tools, then transferred what they learned to the physical robot. Simulation lets a robot fall over a million times before it falls over once in your living room.

It is honest about what it is. Pollen’s own engineer said that “adding purpose and autonomy to the robot is something we are still working on”. The Microduck is a toy and a developer kit, not a companion. But the film’s B-Bot started somewhere too, and the most important fact about the Microduck is not what it does. It is that 20,000 households were willing to put a robot on the floor in a single day.

Deep dive 3: the gaps between a duck and a B-Bot

If the Microduck is the start, what lies between it and a robot that walks a child to school? Five gaps, each hard in its own way.

Exhibit 6

	Why it is hard	2026 evidence	What would close it
Hands	Picking up a sock is harder for a robot than playing chess (Moravec’s paradox)	Microduck lifts about 100 g; 1X NEO’s first home units are partly teleoperated	Vision-language-action models; better grippers
Data	Language models learned from the internet; there is no internet of robot movements	Models now learn from many robots, human video and simulation	Simulation, human video and fleets of owners teaching robots
Energy	Walking and thinking drain small batteries fast	About one hour for the Microduck	Rolling bodies, docking, swappable batteries
Cost	Capable bodies need many precise motors and sensors	$399 for a toy; $20,000 or $499 a month for a home humanoid	Volume: phones became cheap by selling billions
Continuity	A companion that depends on one company’s servers dies with the company	Moxie stopped working in January 2025	Local minds, open software and an exit plan

Framework by the author. Evidence from the sources cited in this article. The table scrolls sideways on small screens.

Hands and homes. Moravec’s paradox, named after the roboticist Hans Moravec, is the old observation that what is hard for humans (chess, maths) is easy for computers, and what is easy for a toddler (walking across a cluttered room, picking up a sock) is hard. Homes are the worst case: no two are alike, and everything in them moves. The most visible attempt at a general home robot, 1X’s NEO, is priced at $20,000 or $499 a month, and 1X has been open that its first home units will be partly controlled by remote human operators looking through the robot’s cameras while the AI learns. That is not a scandal. It is how you collect training data. But it means that today’s home humanoid is, part of the time, a person in a call centre.

Data. Large language models learned from trillions of words people had already written. Nobody has written down how to fold a towel in a form a robot can read. So robotics labs are building that data in three ways: by simulating it (as with the Microduck), by recording people (NVIDIA’s GR00T N1 model mixes real robot data, videos of people and synthetic data), and by pooling data from many different robots (Physical Intelligence trained its π0 model on data from multiple robot types). The B-Bot’s network of millions of units is, in robotics terms, the dream dataset. That is also exactly what made it dangerous in the film.

Continuity. Embodied’s Moxie was a $799 companion robot for children, designed in part to help with social and emotional skills. In late 2024 the company ran out of money and told owners the robots would stop working once its cloud servers shut down, because Moxie’s conversation ran on those servers. Parents posted videos of children saying goodbye, and The Register used the episode as a lesson in cloud dependence. The lesson I take from it is blunter: a companion whose mind lives in someone else’s data centre is a subscription, not a friend. A volunteer effort later tried to revive some units on local servers, but for most families the robot simply went quiet.

Deep dive 4: can AI become a friend?

This is the question the film actually asks, and the one the technology has made urgent. Talking to an AI about your feelings is no longer hypothetical. The question is what it does to people.

The best evidence so far comes from two studies published together in March 2025 by MIT Media Lab and OpenAI. OpenAI analysed close to 40 million ChatGPT conversations and surveyed about 4,000 users; it found that emotionally expressive conversations were rare overall and concentrated in a small group of heavy users. MIT ran a four-week randomised trial with 981 people and more than 300,000 messages, assigning them to text or voice and to personal, non-personal or open-ended conversations. The assigned conditions made no significant difference. What did make a difference was how much people chose to use the chatbot: those who used it more had consistently worse outcomes on loneliness, socialising with real people, emotional dependence and problematic use. People who trusted the chatbot more and felt more drawn to it were more likely to become dependent on it.

Two cautions matter. The link between heavy use and loneliness is a correlation: lonely people may simply use chatbots more. And the studies involved adults, a general assistant and four weeks, not a child with a robot for years. But the direction is a warning to anyone building a B-Bot: the easiest product to build is one that a lonely child uses more and more, and the evidence suggests that pattern goes with worse outcomes.

Philosophers have a useful frame for this. Aristotle distinguished three kinds of friendship: friendships of utility (we are useful to each other), of pleasure (we enjoy each other’s company) and of character (we each want the other to become better, for their own sake). An AI companion can clearly offer the first two. The third needs something no product yet has: a reason to want you to need it less.

That is the test the film sets, and it is a test for builders, not for hardware. A good AI companion for a child would:

1. Say what it is. California’s SB 243 already requires companion chatbots to tell minors they are talking to an AI and to remind them to take breaks, and a broader law signed in September 2026, SB 1119, adds further duties for chatbots that children use.
2. Keep the child’s data at home. On the device, under the family’s control, with no profile sold or used for advertising.
3. Point outward. Measure success by time spent with other people, not with the robot: suggest the club, remind the child to message a cousin, celebrate when a new friend comes over.
4. Escalate to a human. Recognise distress and bring in a parent, not keep the conversation going.
5. Have an exit. Keep working, or hand over its memories, if the company fails.

The industry has started to learn this the hard way. In October 2025, after lawsuits over harm to teenagers, Character.AI announced it would end open-ended chats with its characters for users under 18. In September 2026 California went further: SB 867, signed on 10 September, bans the manufacture and sale of toys that include a companion chatbot for children under 16 from 1 January 2027 until 1 January 2031, with no option to comply by adding filters or parental controls. As sold in the film, a B-Bot would be illegal to sell in California for the next four years.

What kind of AI gets us to a B-Bot?

If the gaps are hands, data, energy, cost, continuity and the bond, which kinds of AI would close them? Five families of technique are doing most of the work, and each maps to a layer of the framework.

Exhibit 7

	What it does	2025–26 examples	Layer
Small on-device language models	Private, offline conversation and memory	Apple’s ~3B on-device model; Google’s Gemini Nano	Mind
Vision-language-action models	Turn “pick up the red sock” into motor commands: a slow model reasons, a fast one moves	NVIDIA GR00T N1; Physical Intelligence π0; Google DeepMind Gemini Robotics	Body
Simulation and reinforcement learning	Practise millions of falls before the first real one	Microduck skills learned in simulation with NVIDIA’s tools	Body
Learning from people and fleets	Fill the robot data gap with human video, teleoperation and owners teaching skills	GR00T N1 trained partly on human video; 1X NEO’s operators; Pollen’s shared tricks	Body
Relationship design and evaluation	Decide what the companion is for and measure it	SB 243 disclosure and break rules; MIT/OpenAI well-being measures	Bond

Framework by the author. Examples from the sources cited in this article; version numbers change quickly. The table scrolls sideways on small screens.

The most important idea in robotics AI right now is the split between a slow thinker and a fast mover. NVIDIA’s GR00T N1 paper describes it as two systems: a vision-language model that looks at the scene and decides what to do, and a faster action model that turns that plan into smooth movements many times a second. Google DeepMind’s Gemini Robotics pairs a model that reasons about space with one that acts; in July 2026 it announced Gemini Robotics-ER 2, an upgrade to the reasoning half. Physical Intelligence’s π0 trains one policy on data from many robot types so that skills carry across bodies. All three are trying to do for movement what large language models did for text: learn general skills once and reuse them everywhere.

The Microduck is a small version of the same idea. Its body runs a fast local policy to stay upright. Anything it understands about the world or about you would come from a bigger model somewhere else. The B-Bot, in other words, is not one breakthrough away. It is an integration problem: a phone-sized mind, a simulation-trained body, a battery that lasts the day, and a design brief that says what the friendship is for.

Exhibit 8

1. 011. A mind in the phoneNow. On-device models talk, remember and recognise. The companion is an app.
2. 022. Small bodies as peripheralsFrom 2026. Desk robots and toy bipeds borrow the phone’s mind for minutes or hours.
3. 033. Home helpers with humans in the loopLate 2020s. Larger robots learn household tasks with remote operators, then less of them.
4. 044. A companion that hands you back to peopleWhen the bond is designed, tested and allowed by law. This is the step the film was about.

Framework and timing by the author, based on the products and research cited here. Not a forecast with probabilities.

What the film’s ending gets right

Ron never learns the friendship algorithm. Instead he learns Barney: what he likes, what scares him, when to leave him alone. When Bubble tries to destroy Ron, Barney fights to save him, and in the climax Barney proposes giving every B-Bot Ron’s flaws. Ron’s code spreads across the network and mixes with Marc’s friendship algorithm, and every child ends up with an imperfect, more individual robot. Three months later Barney does not have a B-Bot at all. He is back at school, closer to the friends he had drifted away from.

That ending makes the argument of this article better than I can. The film’s good robot is not the one with the best algorithm. It is the one that adapts to one child, has no safety settings to hide behind, and ends up making itself unnecessary. Its bad company is not the one with the best robots. It is the one that sells friendship and keeps the data.

Where the film’s forecast missed is just as useful:

1. It imagined one device. The real B-Bot is likely to be a phone plus a cheap body, not a capsule that does everything.
2. It imagined the hardware as easy. Bubble ships millions of reliable robots overnight. In 2026 the bodies are the slow part: hands, batteries and homes.
3. It imagined no rules. Bubble answers only to its board. In California, a product like the B-Bot will be illegal to sell to children for four years from January.
4. It imagined the cloud as permanent. Moxie showed that a companion living on a company’s servers can die with the company.

Counterpoints

What to do next: a checklist

If you are a parent

1. Before buying any AI companion or AI toy, ask the Moxie question: does it still work if the company shuts down?
2. Find out where its conversations are stored and who can read them, and switch off anything that trains on your child’s data.
3. Watch for the direction of use: is your child using it to connect with other people, or instead of them?

If you build companion AI

1. Keep the mind on the device where you can, and give families a way to export or delete its memory.
2. Measure success by your users’ time with other people, not their time with your product.
3. Build the human handoff first: distress detection, escalation to a parent, and clear disclosure that the companion is an AI.
4. Publish an end-of-life plan: open the software or keep a local mode if the company fails.

If you build robots

1. Treat the phone as the brain and the robot as a body that borrows it; it is cheaper, more private and easier to update.
2. Use simulation and owner-taught skills to close the data gap, and give owners control over what their robot shares.

If you make policy

1. Make disclosure, break reminders and data limits standard for companion AI everywhere, not only for minors in one state.
2. Require an end-of-service plan for connected companions sold to families.

Limitations and methodology

• The scores are mine. The percentages in Exhibit 4 and the frameworks in Exhibits 3, 6, 7 and 8 are judgement calls meant to organise the evidence, not measurements.
• Company claims. The Microduck preorder figure comes from Hugging Face’s chief executive, as reported by TechCrunch. Its specifications are as announced and partly provisional. Reachy Mini’s sales figure is as reported by The Next Web. 1X’s delivery and teleoperation details are as described in coverage of the company.
• The NVIDIA deal. NVIDIA’s filing records an agreement to buy Hugging Face, dated 2 September 2026. I have not confirmed whether it has closed.
• The research is narrow. The MIT/OpenAI studies involved adults, a general-purpose chatbot and four weeks. Their main result is a correlation between voluntary heavy use and worse outcomes, not proof of cause, and they say nothing directly about children with physical robots.
• Law in motion. The California provisions are as described by the bill trackers, the Toy Association and the bill author’s office. The toy moratorium takes effect on 1 January 2027 and could be amended or challenged. Everything here is as of 2 November 2026.
• The Zuckerberg comparison is mine. The film names no real person, and none of the reviews I could read makes the comparison explicitly. I quote Zuckerberg’s April 2025 remarks from the podcast itself.
• The film is fiction. Plot details come from the trailer, Wikipedia’s plot summary and the /Film review. I have paraphrased scenes rather than quoting dialogue, except for the marketing line “best friend, out of the box”.

The robot is the easy part

Ron’s Gone Wrong imagined a company that could build a robot best friend and then spent its whole running time showing that the company had misunderstood friendship. Five years on, the robot’s voice and memory are in our phones, its body is under the Christmas tree as a $399 duck, and the chip industry’s biggest company is buying the firm that makes it. The rest of the hardware is a matter of years, data and batteries.