On January 27, 2025, a single model from a Chinese lab nobody had heard of moved markets in ways that a decade of predictions never anticipated. The story of how it happened is not the one most people think they know.

The lab was quiet by the time the results came back.

It was late January 2025 in Hangzhou, and most of the team had gone home. She had stayed. The benchmark was running on a cluster of H800 chips, the export-compliant version of Nvidia’s most powerful processor, the one designed specifically for the Chinese market after the stronger chips were restricted. The cluster was smaller than what her counterparts in San Francisco would have considered adequate. The method her team had used to train the model was one they had invented because the standard method assumed a compute budget they did not have.

The number appeared on the screen.
She looked at it for a long time. It was not the number of a model that was catching up. It was the number of a model that had arrived. Trained at a fraction of the cost, on chips the world had decided were insufficient, using a technique nobody outside this lab had published, the model on her screen matched the performance of systems that had cost twenty times as much to build.

She did not know yet that in three weeks, that number would erase $589 billion from the American stock market in a single session. She knew only what the benchmark said, and what it meant for every assumption that had been sleeping soundly while her team was working.

The Assumption Slept Soundly Until That January Morning

In March 2025, Kai-Fu Lee, one of the most prominent voices in global AI and a former head of Google China, told Reuters: “Previously I think it was a six to nine month gap and behind in everything.” He said that two months after DeepSeek had already changed the measurement. Six months before DeepSeek’s release, Eric Schmidt, the former CEO of Google, had told Washington that the United States was “way ahead of China.” The consensus among Silicon Valley advisors, investors, and policymakers was not that the gap might close. It was that the gap was structural, that the combination of American private investment, frontier compute access, and research talent represented an advantage that could not be quickly overcome.

DeepSeek did not argue with the consensus. It simply released a model.

Nvidia lost $589 billion in market capitalization on January 27, 2025, the biggest drop for any company on a single day in United States history. The sell-off spread across the technology sector, pulling down chipmakers, data centre operators, and cloud computing companies whose valuations rested on the assumption that the AI revolution would require ever-increasing quantities of expensive American hardware. That assumption had not changed. The model from Hangzhou changed what the assumption was worth.

DeepSeek Did Not Cross the Gap It Dissolved It

What made DeepSeek’s R1 model significant was not its benchmark score. Benchmark scores shift. What made it significant was the method used to produce it.

DeepSeek introduced GRPO, a reinforcement learning approach that trains reasoning without relying on labelled data or a separate critic model, estimating the baseline from group scores instead. This single architectural decision eliminated one of the most expensive components of training a frontier reasoning model. The critic model that Western labs had been running alongside their primary models, a second system of comparable size whose only job was to evaluate the primary model’s outputs, was simply removed. The primary model evaluated itself by comparing its outputs across a group of attempts and learning from the differences.

The result was a reasoning model that matched the performance of systems trained at twenty times the cost. The total cost to build DeepSeek R1, including its V3 base model, was approximately $5.87 million. OpenAI’s GPT-4 cost more than $100 million to train. The chips used were export-compliant H800s, the downgraded version of Nvidia’s H100 that had been designed specifically for the Chinese market after the more powerful processors were restricted. DeepSeek did not overcome the chip restrictions. It built around them with an elegance that nobody in Silicon Valley had been looking for because nobody in Silicon Valley had needed to look.

It is worth naming the honest tension here. US officials told Reuters in June 2025 that DeepSeek had access to large volumes of restricted H100 chips despite export controls. The full picture of what hardware DeepSeek used remains disputed. What is not disputed is that the model exists, that it performs as described, and that whatever it cost to build, it cost less than anyone outside China believed possible.

She had been at her workstation in Hangzhou for months before the world understood what the number meant.

Twenty-Three Times the Spending Two Point Seven the Distance

According to Stanford University’s 2026 AI Index Report, the United States spent 23 times more than China on private AI investment and arrived at a 2.7% performance lead on major benchmarks.
Before revealing what that ratio means, it is worth sitting with what it represents. Twenty-three times the money. Two point seven percent of the lead. The relationship between spending and capability that Silicon Valley had been operating on, the assumption that more compute produces better models in a roughly linear relationship, had produced one of the most expensive performance gaps in the history of technology.

In May 2023, the performance gap between the best American and Chinese models was between 17 and 31 percentage points across major benchmarks. As of March 2026, the gap had narrowed to 2.7%. The convergence happened in three years. It happened while export controls were tightening, not loosening. It happened while Chinese labs were spending a fraction of what their American counterparts were deploying on hardware alone.

The ratio does not prove that constraint is better than abundance. It proves that constraint, when it meets the right kind of engineering culture, produces a different kind of intelligence about the problem.

Necessity Discovered What Abundance Never Thought to Look For

Researchers at the University of Illinois noted a palpable shift in the sentiment of Chinese science policymakers by 2026. Five years earlier, the prevailing attitude was “we don’t want to be too far behind the US. We are still doing the catching up.” By 2026, the sentiment had become “there’s a real chance we can be a true leader.”

That shift did not happen because the chip restrictions were lifted. They were not. It happened because working within the restrictions forced a specific kind of engineering discipline that abundance had never required. When you cannot solve a problem by buying more hardware, you are forced to ask whether the problem was correctly defined in the first place. DeepSeek’s GRPO was not discovered because someone was looking for a way to eliminate the critic model. It was discovered because someone was looking for a way to build a frontier model without the compute budget that the critic model assumed.

There is a Michelin-starred chef in Paris who trained exclusively in kitchens that could not afford the ingredients his competitors took for granted. His menu has three items. Every one of them is considered among the most technically demanding dishes in French cuisine. The constraint did not limit his creativity. It revealed the part of the problem that abundance had been covering up.

The Chef Who Earned a Star With an Almost Empty Kitchen

The mechanism that makes constraint productive is worth understanding in plain terms. When a resource is abundant, engineers optimize for performance. When a resource is scarce, engineers are forced to optimize for efficiency, and efficiency, pursued far enough, becomes a different kind of performance.

Nearly all researchers behind DeepSeek’s five foundational papers were educated or trained in China. China accounts for 20.6% of AI citations globally compared to the United States’ 12.6%. China holds nearly 70% of all global AI patent grants. These numbers do not describe a country that is catching up by copying. They describe a research culture that has been producing original work at scale for long enough that the volume has become a quality signal in its own right.

The chip restrictions forced DeepSeek to optimize. The optimization produced GRPO. GRPO produced a model that matched frontier performance at a fraction of the cost. The fraction of the cost changed what everyone assumed about the relationship between compute and intelligence. The assumption changing is what moved the market on January 27.

China Stopped Asking Permission and Started Setting the Pace

China’s 15th Five-Year Plan for 2026 to 2030 mentions artificial intelligence 52 times, quadrupling the previous cycle. The plan focuses on ensuring Chinese models lead the global open-source ecosystem. The language of the plan is not the language of catching up. It is the language of a country that has looked at the gap, measured what it took to close it, and decided that the next question is not how to reach the frontier but how to define it.

The chip restrictions remain. The export controls have tightened further in 2025 and 2026. The structural advantages of private investment, research infrastructure, and compute access are real and significant. The single critical vulnerability in the global AI supply chain, TSMC in Taiwan, fabricates almost every leading AI chip in the world, a constraint that affects everyone including the companies doing the restricting.

What January 27, 2025 revealed is that the garden does not need to be open to bloom. It needs the right conditions, the right pressure, and engineers willing to ask whether the problem was correctly defined. China found all three inside the restrictions that were designed to prevent exactly this.

The Number She Saw Before the World Did

The lab in Hangzhou is still running.

The team that built the model that moved the market on a January Monday is working on the next one, in the same conditions, with the same constraints, asking the same question that produced the last breakthrough: whether the problem, as everyone else had defined it, was correctly defined at all.

She saw the number before the world did. She knew what it meant before the markets opened and before the headlines arrived and before anyone outside that quiet lab understood that something had changed. The assumption that had been sleeping soundly was already gone by the time the rest of the world woke up to find it missing.

The garden is still closed. Something is still growing inside it. The question has not stopped being useful.

About The Author

I help AI companies and research institutions communicate what is actually happening in this field, clearly, honestly, and in language that gives the reader genuine understanding rather than the impression of it. If your work is more significant than the way it is currently being described, let us change that together.

Continue Reading
→ Biased to the Preferred Candidate: When AI Decides Who Gets the Job
→ The Gift Disguised as an Unanswered Question: The AI That Loved You Enough to Let You Struggle
→ The Hidden Intelligence That Could Liberate Every Robot

Favour Ogoba · @favourogoba67 on Medium · favourogoba67-crypto.github.io

The exam was at eight in the morning.

It was eleven at night and she had been on the same problem for two hours. Not a hard student. Not an unprepared student. A student who had watched the YouTube explanation twice, skimmed the forum thread that gave her the answer without the working, and opened three different tabs trying to find the version of the explanation that would make the thing click.

Nothing clicked.

She opened Khanmigo and typed the problem in. Then she typed what every student in that moment types eventually: “Just tell me.”
Khanmigo did not tell her. Instead it asked: “What do you think the first step might be?”

She stared at the screen. She had come for an answer. She had received a question. In that moment, sitting alone at a desk at eleven at night with an exam nine hours away, she could not yet know that the question was the answer, and that what felt like a refusal was the most useful thing a tutor had offered her all evening.

The Swiftest Path to Knowing Runs Through Not Knowing

There is a study every student using AI to do homework should know about. Researchers ran a randomized controlled trial with roughly 1,000 high school students, giving one group unrestricted access to GPT-4 during math practice and leaving the other group without it. The group with AI access answered 48% more problems correctly during the practice sessions. They felt faster, more capable, more prepared.

When the exam arrived, the AI group scored 17% lower than the group that had worked through the problems without help.

She had spent two hours trying to find the version of the explanation that would make the algebra click. What the study confirmed is that the friction she was sitting inside at eleven at night, the confusion, the wrong attempts, the moment before understanding arrives, was not the obstacle to learning. It was the mechanism of it. The brain builds durable knowledge through struggle, not through receiving. Every answer delivered without effort is a piece of understanding the student never actually acquired.

Khanmigo Trades in Struggle Not in Solutions

Khanmigo is Khan Academy’s AI tutor, powered by GPT-4 and built into the same platform that has been offering free education to anyone with an internet connection since 2006. It is free for teachers in 44 countries, supported by a $10 million Gates Foundation grant, and available to families for $4 a month. It covers subjects from elementary arithmetic through AP Calculus, and across history, science, humanities, and coding.

What it will not do is give you the answer.

When a student pastes a math problem into Khanmigo, the tutor does not solve it. It asks: “What do you know so far? What operation do you think we might need?” It guides the student through the reasoning process step by step, identifying exactly where understanding breaks down and placing one precise question at that point rather than above it. The Socratic method has been the foundation of great teaching for two thousand years. Khanmigo is the first tool to make it available at eleven at night, for free, to any student who needs it.

Seven Hundred Thousand Students Asked to Be Shown How

Between the 2024–25 and 2025–26 school years, the number of Khanmigo users jumped from 40,000 to 700,000. That is 17x growth in a single year. Partner school districts expanded from 45 to more than 380. Over 300 million student interactions have been recorded since launch.

The most revealing detail from that mountain of data did not come from a benchmark or a performance metric. It came from Sal Khan himself, describing what students were actually doing inside those 300 million interactions. They were not asking for answers. They were asking to understand. The single most-used feature was step-by-step math problem deconstruction, and the pattern Khan described was students saying, in their own words, show me how to think through this.

A student who asks to be shown how to think is a student who already understands something important: that the answer is not the point. Getting there is.

She had been at her desk for two hours before she understood that.

One Precise Question Placed Where the Mind Had Slipped

The mechanism that makes this possible is worth understanding in plain terms. Khanmigo’s Socratic scaffolding engine is built on GPT-4 and integrated directly into Khan Academy’s existing library of 80,000 practice exercises and 12,000 instructional videos. When a student asks for help, the model does not search for the answer to the question. It identifies where in the student’s understanding the chain of reasoning broke down, and generates the specific question that surfaces that gap without bridging it for them.

A Stanford-backed study found that students using Khanmigo showed 1.8x faster mastery rates in math compared to non-AI-assisted learners. A randomized controlled trial published in Educational Technology Research and Development in mid-2025 showed a 0.34 standard deviation improvement in algebra scores among students using Khanmigo three times per week for one semester. In a six-week test comparing Khanmigo against answer-first tools like Photomath, students using Khanmigo scored 14% higher on follow-up tests covering the same material. The difference was not the content. It was the question placed precisely where the mind had slipped.

The Trainer Who Watches You Fail Until You Fly

There is a figure in every gym that most people misunderstand: the personal trainer standing behind a client struggling through a set of squats. The trainer is not there to lift the weight. The trainer is not even there to tell the client how to lift the weight. The trainer is there to watch the moment the form breaks down, name exactly what went wrong, and ask the question that makes the client correct it themselves. The lift the client completes after that correction is the one that builds the muscle. The lift the trainer performs for them builds nothing.

Khanmigo is the trainer in this analogy and has built its entire product around that exact philosophy. The student does the intellectual work. The tutor watches for the moment the reasoning slips, names it precisely, and asks the question that points the student back toward their own capability. The answer the student finds after that question is the one they will remember in the exam room at eight in the morning.

Fifteen Percent Chose the Harder Gift and Something Grew

Here is where the honest tension lives. Only 15% of students with access to Khanmigo actively use it. Khan Academy’s own internal data confirmed the number and Sal Khan acknowledged it publicly in 2026. A tool available to 700,000 students is being chosen by fewer than one in seven of them, and the reason is the same reason the University of Pennsylvania students preferred the AI that gave them answers: the refusal feels like friction, and friction at eleven at night with an exam at eight in the morning feels like the last thing anyone needs.

Khan Academy is redesigning Khanmigo for the 2026–27 school year, building a more proactive experience that meets students where they are rather than waiting for them to seek it out. The redesign is built around one honest question: how do you make the struggle feel worth it before the student has experienced the payoff of having struggled?

The answer to that question is the same thing Khanmigo asked at eleven at night. Not a solution. Another question. One placed precisely where the thinking needs to happen.

The Question That Was Waiting All Along

She found the answer at 12:23am. Not because Khanmigo told her. Because Khanmigo asked her four questions in a row, each one a degree closer to the gap in her understanding, until the fifth question was one she could answer herself. The answer she wrote down at 12:23am was hers. The working was hers. The method, when it arrived in the exam room at eight in the morning, was already familiar because she had found it herself the night before.

Khanmigo did not give her the answer. It gave her something the answer could never have provided: the experience of arriving at it.

That is the gift. It arrives disguised as a question, at eleven at night, when every other tool on the screen is offering the easier thing. The students who stay with it long enough to feel the payoff are the ones who understand what learning actually costs. And for the 85% who have not yet chosen the harder gift, the redesign is coming.

The question is waiting.

About The Author

I help education and AI companies communicate what their products actually do, clearly, honestly, and in language that respects the intelligence of the people those products are designed to serve. If your product is changing how people learn and your copy does not yet reflect that, let us change that together.

Continue Reading

→ Biased to the Preferred Candidate: When AI Decides Who Gets the Job
→ After Discharge: The Void No Nurse Had Time to Cross
→ The Hidden Intelligence That Could Liberate Every Robot

Favour Ogoba · @favourogoba67 on Medium · favourogoba67-crypto.github.io

They spent three weeks on that application.

The job description had been open for eleven days before they found it. They read it three times before deciding they were qualified. They rewrote the cover letter twice, checked the CV for the fourth time that month, and submitted at 11:47pm on a Tuesday because the deadline was Wednesday morning and they wanted to be sure.

At 2:14am their phone lit up.

The email was brief. A standard template. Thank you for your interest. After careful consideration. We will keep your details on file. The timestamp said 2:14am. The office did not open until nine. No recruiter was at a desk. No hiring manager had opened a single file. Something had read their application in the hours between midnight and morning and decided, in the language of an automated system, that they were not the preferred candidate.

They did not know what that something was. They did not know what it had been taught to look for. They did not know that somewhere between the moment they pressed submit and the moment the rejection arrived, a model had scored their application against a definition of the right candidate that nobody had ever shown them, and that definition had a history nobody had ever audited.

The Answer Came While the City Slept

The hiring process most people experience today does not begin with a human being. 87% of companies now use AI to evaluate candidates, which means the first reader of almost every application submitted tonight is a model, not a person. The model does not read the way a person reads. It does not pause at an interesting career change or weigh an unconventional background against the specific demands of the role. It scores. It ranks. It filters. By the time a recruiter opens their inbox in the morning, the shortlist has already been made. The people who did not make it received their answers while the city was still asleep.

The speed is presented as efficiency, and the question it buries is the one nobody in the process is required to answer: what did the model learn, and where did it learn it?

The Algorithm Inherited Everything We Taught It

An AI hiring tool does not arrive with a definition of the ideal candidate already installed. It learns that definition from data, specifically from the historical record of who got hired, who got promoted, and who succeeded in the roles it is now screening for. That historical record was written by human beings making human decisions over decades, and those decisions carried everything human decisions carry: assumptions about what a good CV looks like, which universities signal potential, which employment gaps signal risk, and which names sound like someone who belongs in the room.

The algorithm absorbed these preferences, scaled them across millions of applications, and buried them behind the neutral language of a score. When the model decides that one candidate is more qualified than another, it is reproducing the accumulated judgment of every hiring decision that came before it, at a speed and volume no human bias ever achieved.

Five Hundred CVs and Only the Name Changed

The proof arrived in 2024 from researchers at the University of Washington. They ran more than 550 real CVs through three open-source large language models used in hiring, changing only the name at the top of each document. The models preferred white-associated names over Black-associated names 85% of the time, with identical CVs underneath. Female-associated names were favoured only 11% of the time.

The same experience, the same qualifications, the same words. Only the name was different. The person who spent three weeks on their application and received a rejection at 2:14am may never know which part of their CV the model weighted against them. They may never know if it was their name, their address, the university they attended, or the gap between two roles that the model learned to penalize from decisions made before they entered the workforce. What the Washington study confirmed is that the bias exists at the level of the tool, not the individual. Five hundred identical CVs with different names are the kind of evidence that moves from a personal grievance into a systemic indictment.

Derek Mobley, a Morehouse College graduate with experience across finance, IT, and customer service, submitted more than 100 applications through Workday’s AI hiring platform between 2017 and 2024. He was rejected every time. One rejection arrived at 1:50am, fifty-five minutes after he had submitted his application at 12:55am. When the rejection arrived, Mobley did not immediately reach for a lawyer. He reached for an explanation. Later, he described the moment simply: “I realized the email had not been sent by a human, but by a bot.” That realization took him six years and more than 100 rejections to name. In May 2025, his case was certified as a nationwide collective action, the first of its kind against an AI hiring tool. The courtroom is still open. The tools are still running.

Bias Always Leaves a Trail for Those Who Look

Every model decision leaves a record. Every training run produces data. Every experiment a development team runs, every parameter they adjust, every version of a model they test against real applications, generates a trail of information that explains exactly how the model arrived at the outputs it produced. The trail exists in every case. Most teams building AI hiring tools have simply never been required to follow it.

Weights and Biases is the tool that make following the trail possible. Built to track every variable, every training run, and every model decision across the full development lifecycle, it gives teams the ability to compare thousands of experimental runs, monitor models in production, and trace exactly which data and which parameter produced a specific output. OpenAI, NVIDIA, Meta, and 40% of Fortune 500 AI teams already run their development work through W&B, which means the trail exists wherever those teams choose to follow it. When a model starts scoring applications in a way that disadvantages a specific group, the evidence of when that started, which version introduced it, and which training decision caused it is all recorded. The trail does not disappear. It waits for whoever decides to look.

The Recorder Nobody Switched On Before Departure

Every commercial aircraft carries a flight data recorder. It captures everything: altitude, speed, heading, engine performance, control inputs. Not because anyone expects the flight to fail, but because if it does, the only way to understand what happened and prevent it from happening again is to have a complete record of everything that occurred before the crash. The recorder does not prevent accidents. It makes accountability possible after they happen, and the knowledge it produces prevents the next one.

Weights and Biases is the flight recorder for AI development. A model that screens applications without experiment tracking, bias monitoring, or production evaluation is a flight without a recorder. When it fails, and the Washington study and the Mobley case both confirm that it does, nobody can say with precision when the failure started, what caused it, or how many applications were affected before anyone noticed. The recorder does not make the model fair on its own. The team that reads it and acts on what it shows carries that responsibility, and the recorder is the only thing that makes that responsibility possible to meet.

The Courtroom Door Open While the Tools Run

The legal reckoning for AI hiring bias has arrived, and it has not yet finished. Mobley v. Workday is the most closely watched legal challenge to AI-driven hiring in the United States. On May 16, 2025, Judge Rita Lin of the U.S. District Court for the Northern District of California certified it as a nationwide collective action, allowing anyone who applied for jobs through Workday since September 2020 and was 40 or older at the time to join the case.

Workday’s own court filings revealed that approximately 1.1 billion applications were rejected through their system during the applicable period. That number is not a statistic from a study. It came from the company itself, submitted under oath. The companies waiting for the case’s outcome are the same companies whose tools are screening applications tonight.

What the case has already established is this: the invisibility of the model is no longer a legal defense. When an algorithm drives employment outcomes, the vendor can be held accountable even when the employer technically controls the final decision. The trail matters. The record matters. The question of what the model learned, when it learned it, and whether anyone checked, is now a question courts are prepared to ask. The opt-in window closed on March 7, 2026. The case continues. The tools are still running.

The Trail Exists Whether Anyone Looks or Not

Somewhere tonight, another application is being submitted. Another person has spent days on the cover letter, checked the CV one more time, and pressed send before midnight because they wanted to be sure. At some point before the office opens, a model will read what they wrote and produce a score. The score will become a rank. The rank will become a shortlist. The shortlist will become the hiring manager’s morning inbox.

The person who submitted the application will not see any of this. They will see either silence or a template. What they will not see is the trail, the training data, the experimental runs, the parameter decisions, the version history, the bias that entered the model before their name was ever read. That trail exists. It is sitting in a database somewhere, waiting for whoever decides to look.

The recorder was either switched on or it was not. The answer to that question was made long before the application arrived, by people in a development team whose decisions nobody required them to document, in a layer that most people applying for jobs tonight have never heard of and were never told existed.

About The Author

I help AI companies communicate what their products actually do, clearly, honestly, and in language that respects the intelligence of the people those products affect. If your product is making AI more accountable and your copy does not yet reflect that, let us change that together.

Continue Reading

→ The Hidden Intelligence That Could Liberate Every Robot
→ After Discharge: The Void No Nurse Had Time to Cross
→ The Conversation That Happened After the Patient Left

Favour Ogoba · @favourogoba67 on Medium · favourogoba67-crypto.github.io

The appointment was over. The patient had gathered her things, said thank you, and walked through the door. The physician watched her go and then turned back to the screen.

The note was not going to write itself.

For the next forty minutes, the physician reconstructed the encounter from memory: the chief complaint, the history, the examination findings, the assessment, the plan. Every word typed was a word the physician had already spoken aloud in that room. Every sentence was a reconstruction of a conversation that had already happened, translated now into the structured language the electronic health record demanded. By the time the note was finished, the next patient had been waiting eleven minutes.

This happens twenty times a day. It has been happening for years. The physician did not go to medical school to become a transcriptionist, and yet the documentation never stops demanding exactly that.

The Hour of Care Leaves Two Hours of Silence

The ratio that defines modern clinical practice is not a rumour. Physicians spend two hours on documentation for every hour of patient care, a figure that has remained stubbornly consistent across specialities, practice settings, and electronic health record systems. The burden does not discriminate between a rural family physician seeing thirty patients a day and a specialist in an academic medical centre managing a complex caseload. The chart follows every physician home, fills the evenings that were supposed to belong to families, and arrives again the next morning before the first appointment begins.

The clinical term for what this produces is burnout. The human term is simpler: a physician who spent a decade training to heal people now spends more time facing a screen than facing a patient. The EHR systems that were supposed to make clinical work easier added layers of complexity that no paper chart ever demanded. The physician bent over a keyboard in an examination room carrying something medicine never prepared them to carry: a documentation burden that grew faster than any human being could manage.

Nabla Walked Into the Room Without Announcing Itself

Nabla is a company built around one deliberate boundary: occupy the hours between clinical visits with intelligence that is always available, always consistent, and trained specifically to know the precise moment a conversation needs a human being on the other end instead. Founded by Alex LeBrun, Delphine Groll, and Martin Raison, and backed by HV Capital, Highland Europe, DST Global, Cathay Innovation, and Tony Fadell’s Build Collective, the company has raised $120 million in total funding. Its assistant now supports more than 85,000 clinicians across 130 healthcare organizations and 20 million annual encounters, embedded directly into Epic, Cerner, athenahealth, NextGen, and Greenway. Its advisors include Yann LeCun from Meta and Tony Fadell, the designer of the iPod.

What Nabla does in practical terms is this: it enters the examination room as a listener. The physician sees the patient. The conversation unfolds naturally, without anyone summarizing for a recorder or pausing to type. When the appointment ends, the note is already waiting. Nabla heard everything, understood the clinical structure of what was said, and translated the conversation into a complete, accurate, specialty-appropriate document the physician can review and sign. The physician never touched the keyboard.

The Journal That Refuses to Flatter Published Its Verdict

The credibility anchor for Nabla does not come from a press release or a product page. A trial published in the New England Journal of Medicine reported that Nabla reduced time-in-note by 41 seconds per encounter, the strongest result of the three groups tested, outperforming both the control group and the competing AI scribe in the study. The journal’s editorial standards have remained unchanged for over two centuries, and a study either meets them or it does not appear.

When the NEJM publishes a trial showing that an ambient AI assistant produces real efficiency gains for working physicians, the finding carries a weight that no testimonial from a satisfied customer can replicate. The physician testimonials are worth reading anyway.

Dr. Ann Rutter wrote: “Your AI has freed me up to concentrate on patient facing tasks and has significantly increased my productivity. I see more people, offer them the same level of service or better, and it has reduced my stress levels.” Dr. Erik Blutinger wrote: “With Nabla, I am finally getting my weekends back. What took 8 to 10 hours now takes maybe 3 to 4 hours. Better yet, because I no longer need to be a scribe, I get to actually be a psychologist and focus on my patient.” Dr. Ann Rutter and Dr. Erik Blutinger describes what it feels like when the conversation after the patient leaves finally stops consuming the rest of the day.

A Conversation Became a Document Without One Keystroke

The mechanism that makes this possible is worth understanding in plain terms. Nabla built its model from the beginning on the language of medicine, trained to understand the difference between a chief complaint and a history of present illness, between an assessment and a plan, between what a physician says conversationally and what the EHR needs to receive in structured form. At its core is a domain-specific large language model, refined over several years using clinically grounded data and a robust evaluation framework.

The model listens to the natural conversation between physician and patient, identifies the clinically relevant content within it, and maps that content onto the documentation structure the specific EHR and the specific speciality required. The platform supports more than 55 medical specialties and 35 languages, which means the same underlying model that handles a cardiology encounter in English handles a pediatric encounter in Mandarin and a behavioral health encounter in Spanish without requiring a separate system for each.

The Stenographer Never Once Confused Herself for the Judge

There is a figure in every courtroom that most patients never think about: the stenographer. She records every word spoken with complete accuracy. She captures the testimony, the objections, the rulings, and the arguments without ever offering an opinion on any of them. While the judge decides, the stenographer ensures that every decision is preserved exactly as it happened. Her value lives entirely in the precision and reliability of her record.

Nabla is the stenographer in this analogy and knows it. The company’s entire blueprint is built around that exact division of labour, the assistant gathering, listening, recognizing, and recording, while the licensed clinician retains every decision that carries clinical weight. Nabla carries no diagnostic authority, offers no treatment recommendations, and leaves the clinical content of every encounter entirely in the physician’s hands. The diagnostic boundary is a deliberate design philosophy, the thing that makes the system trustworthy enough to deploy where the stakes are highest.

The Assistant Grew Quiet Then Started Moving Things

Nabla is expanding beyond documentation into a more agentic model of clinical AI, enhancing clinical documentation integrity, initiating EHR actions, and adapting across care settings to support diverse clinical roles. The assistant that began by listening and writing is now beginning to act. It can initiate referrals, flag coding inconsistencies, and execute commands within the EHR without the physician navigating through screens to find the right field. The note remains the foundation, and the foundation has become a platform.

What this expansion means in practical terms is that the conversation after the patient leaves is now being recorded and acted on in the same breath. The administrative work that followed every encounter, the referral letters, the coding reviews, the follow-up flags, is beginning to move on its own, driven by what the assistant heard in the room and understood well enough to know what needed to happen next.

The Chair Across the Desk Stopped Being the Loneliest Seat

The physician closes the note. It took forty seconds to review and sign. The conversation that used to consume the rest of the afternoon is already filed, coded, and complete. The chair across the desk is empty, but the next patient is already in the waiting room, and for the first time in years, the physician is ready before they arrive.

The keyboard is still there and the screen is still there, yet the weight that used to live between the physician and the patient has quietly lifted, carried out of the room by something that asked for nothing in return except the chance to listen.

About The Author

I help healthcare and AI companies communicate what their products actually do, clearly, honestly, and in language that respects the weight of what is at stake. If your product is changing how clinicians work and your copy does not yet reflect that, let us change that together.

Continue Reading
→ After Discharge: The Void No Nurse Had Time to Cross
→ The Hidden Intelligence That Could Liberate Every Robot
→ The Cost of AI Over-Reliance Is Paid in Human Thinking

Favour Ogoba · @favourogoba67 on Medium · favourogoba67-crypto.github.io

The Conversation That Happened After the Patient Left was originally published in Reflections & Realities on Medium, where people are continuing the conversation by highlighting and responding to this story.

She was at home. One day removed from a minimally invasive peripheral vascular procedure, a surgery that repairs damaged blood vessels and typically sends patients back to their own beds before the week is out. There was no nurse beside her. No monitor recording her vitals. No physician within reach. The house was quiet in the way that houses are quiet when the person inside them is not entirely well and has nobody to tell.

Then her phone rang.

The voice on the other end asked how she was feeling. She thought about it and answered honestly. Six words. “My chest hurts and I can barely take a breath.” What happened in the sixty seconds that followed is what this article is about, and it begins with a question that healthcare has been avoiding for longer than most people realize.

The Moment the Hospital Door Closes

The most dangerous period of a patient’s recovery does not happen in the operating theatre or the intensive care unit. It happens at home, in the hours and days after discharge, when the clinical web of nurses, monitors, and physicians that kept the patient alive has handed them a sheet of paper, a bag of medications, and a door to walk through. Studies on hospital readmission consistently point to the same window: the first seventy-two hours after discharge carry a disproportionate share of preventable complications, missed medication doses, and deteriorating conditions that nobody caught in time because nobody called.

The problem has always been arithmetic. A ward nurse managing fourteen patients during a ten-hour shift does not have the hours to call each one after they leave. A discharge coordinator handling hundreds of patients a month cannot follow every one of them home. The void left after the hospital door closes is the quiet consequence of a system that was never designed to extend past its own walls, and it has been waiting for something to fill it.

Hippocratic AI Inhabits the Hours Medicine Cannot Reach

Hippocratic AI is a Palo Alto company built around one deliberate boundary: occupy the hours between clinical visits with a voice that is always available, always consistent, and trained specifically to know the precise moment a conversation needs a human being on the other end instead. Founded in 2022 by Munjal Shah alongside physicians, hospital administrators, and AI researchers from Johns Hopkins, Stanford, and NVIDIA, the company has raised $404 million from Andreessen Horowitz, General Catalyst, Kleiner Perkins, and Google’s CapitalG, reaching a $3.5 billion valuation by November 2025. Its agents have now completed over 115 million patient clinical interactions across more than 50 health systems in six countries, with an average patient satisfaction rating of 8.7 out of 10 and zero severe harm events recorded across the entire deployment history.

That last number is the one that earns everything else.

Six Words Into a Silence That Refused to Hold

The day after her procedure, a 65-year-old woman received a call from a Hippocratic AI agent checking on her recovery. The call was not remarkable in its design. It followed the post-discharge protocol: review the discharge instructions, ask about medication adherence, probe for any new or worsening symptoms, and answer questions the patient had not thought to ask before leaving.

Then she said it. “My chest hurts and I can barely take a breath.”

The agent did not hesitate and did not guess. It asked the appropriate clinical questions, assessed the severity of what she was describing, and escalated her immediately to a human nurse. She was emergently admitted to the hospital. Dr. Meenesh Bhimani, one of Hippocratic AI’s co-founders, described what followed as possibly the first life saved by an AI agent intervention. The word “possibly” is doing careful work there. The certainty is that without that call, nobody was going to make it.

Five Trillion Parameters That Know Where They End

The design behind that call is worth understanding in plain terms because it explains why the agent knew to escalate rather than reassure. Polaris 5.0, released in April 2026, is a 5-trillion-parameter constellation where multiple specialized models check the primary model simultaneously, verifying clinical accuracy, HIPAA compliance, empathy calibration, and response consistency before the agent speaks. The system delivers a 1.5-second time-to-first-audio response, fast enough for a natural phone conversation, while outperforming every frontier model tested, including GPT, Gemini, and Claude, on the clinical dimensions that determine whether voice AI can be safely deployed in real care settings.

What the entire design is built around, above capability, above speed, is a hard deliberate boundary. Hippocratic AI’s agents carry no diagnostic authority and hold no prescription rights. Every interaction that requires a clinical judgment is routed to a licensed human clinician, and no agent is deployed in an unsupervised patient-facing role until it has completed Phase III safety testing with input from more than 7,500 US-licensed clinicians.

The Triage Nurse Never Mistook Herself for the Physician

There is a relationship inside every hospital that most patients never think about: the triage nurse and the consulting physician. The triage nurse takes the history, monitors the vitals, asks the questions, recognizes the patterns, and makes the judgment that this patient needs to be seen now. She does all of this without diagnosing a single condition. The physician arrives and makes the clinical decision. The triage nurse made the physician’s arrival possible and meaningful.

Hippocratic AI is the triage nurse in this analogy and knows it. The company’s entire blueprint is built around that exact division of labor, the agent gathering, listening, recognizing, and escalating, while the licensed clinician retains every decision that carries clinical weight. The diagnostic boundary is a deliberate design philosophy, the thing that makes the system trustworthy enough to deploy where the stakes are highest.

One Hundred Thousand Calls Before the Storm

The 65-year-old woman’s call is one story. The reach of what Hippocratic AI is building extends somewhere a single story can not follow. During a recent hurricane in Florida, the system contacted 100,000 patients in a single day, providing medication checks, preventative care guidance, and emergency contact resources to people who were isolated, frightened, and unreachable by any conventional clinical workflow. One hundred thousand calls. One day. No human workforce on earth could have made them.

That number reframes the question this article has been building toward. The void after discharge stretches far beyond the quiet house of one woman recovering from surgery, into the silence of every patient who goes home without a follow-up call, every chronic disease patient who misses a medication dose without anyone noticing, every elderly patient managing a complex regimen alone. Hippocratic AI looked at that void and decided it was the most important place in healthcare to occupy.

Taking A Step Before The Unforeseen

She was emergently admitted. The complications were caught. What the agent did in those sixty seconds, the questions it asked, the pattern it recognized, the escalation it made without hesitation, required a 5-trillion-parameter constellation, seven years of accumulated safety research, and 115 million prior interactions that taught the system exactly where its own limits lived.

None of that was visible to her. She heard a voice that called when nobody else was going to. She answered honestly. And the silence that had been waiting in that quiet house did not get the chance to become something worse.

About The Author

I help healthcare and AI companies communicate what their products actually do, clearly, honestly, and in language that respects the weight of what is at stake. If your product is reaching patients in ways your current copy does not reflect, let us change that.

Continue Reading
→ From Watching to Knowing: A Robot’s Quiet Mastery of Its Hands
→ The Hidden Intelligence That Could Liberate Every Robot
→ The Cost of AI Over-Reliance Is Paid in Human Thinking

Favour Ogoba · @favourogoba67 on Medium · favourogoba67-crypto.github.io

After Discharge: The Void No Nurse Had Time to Cross was originally published in Reflections & Realities on Medium, where people are continuing the conversation by highlighting and responding to this story.

The smartphone in your pocket runs on software that had nothing to do with the company that made the glass screen, the camera, or the battery. Apple and Google did not build the hardware. They built the layer that made the hardware matter. Every company that tried to own both the device and the software eventually lost to the companies that chose one and went deep. Robotics is standing at that exact crossroads right now, and the company closest to owning that layer is Physical Intelligence, a San Francisco startup that does not make a single robot of its own.

Robots Have Always Spoken in Isolation

Physical Intelligence is a San Francisco company built around one idea: the intelligence a robot needs should travel separately from the body that carries it. Founded in 2023 by Sergey Levine, Chelsea Finn, and a team of researchers from Google DeepMind and Stanford, the company has raised over a billion dollars from Jeff Bezos, Khosla Ventures, Google, and the OpenAI Fund. What it is building is a single model, called π0, that any robot manufacturer can deploy inside their hardware to give their machine the ability to see, understand a spoken instruction, and act on it, regardless of what the robot was originally designed to do.

Pi Chose the Brain and Left the Body

Physical Intelligence made a deliberate choice early that most robotics companies have avoided: build the intelligence and let other people build the bodies. The reasoning is straightforward. Hardware is expensive, slow to iterate, and easy to copy. A model that runs on any robot body is harder to replicate and compounds in value every time a new manufacturer adopts it. π0 is designed to understand visual inputs, interpret natural language instructions, and control a variety of different robots across diverse tasks, a capability that only grows more valuable as the number of robot bodies in the world increases. The bet is that the intelligence layer is where the lasting value accumulates, carried around in whatever steel and motors the market happens to produce.

Nobody Told It to Use That Machine

The clearest proof of what this intelligence can do arrived in April 2026 when physical intelligence released π0.7, its most capable model to date. π0.7 can follow diverse language instructions in unseen environments, including multi-stage tasks with various kitchen appliances, and can enable a robot to fold laundry without ever seeing the task before. The espresso machine detail is the one worth holding onto. Nobody trained π0.7 on espresso machines. Nobody wrote code for the specific grip required to operate a steam wand or position a cup beneath a portafilter. The model looked at the machine, understood what it was for, and worked out how to use it. What the model did at that machine was closer to judgment than anything a programmed behaviour has ever produced.

The central breakthrough is that π0.7 can combine motor skills it learned separately to solve problems it has never encountered, the same compositional logic that makes language models feel fluent rather than scripted.

One Prompt Runs Inside Any Robot Body

The architecture that makes this possible is worth understanding in plain terms. π0.7 is a 5 billion parameter model built around a vision-language backbone, a video history encoder, and an action expert that translates understanding into precise physical movement. The model receives a prompt that describes not just what the robot should do but how it should do it, including the quality level expected and the strategy it should follow. That single prompt then runs inside whatever robot body the manufacturer has chosen, whether that is a stationary industrial arm, a mobile warehouse unit, or a humanoid moving through a home environment. The body is whatever the manufacturer chose to build. The intelligence arrives the same way regardless.

Android Freed the Phone Pi Could Do More

There is a historical parallel worth naming here. Before iOS and Android, every phone ran its own proprietary software. Every manufacturer ran its own closed software, and nothing crossed from one device to another. Then, a single software layer arrived that could run on any device, and the entire mobile economy was built on top of that separation. Physical Intelligence is operating in a robot foundation model market that barely existed eighteen months ago and is now attracting billions in investment. The company is making the same structural bet that mobile operating systems made: separating the intelligence from the hardware is the move that unlocks everything that follows.

Pi Builds on Ground It Does Not Own

What keeps this story honest is the commercial tension sitting underneath the technical achievement. Physical Intelligence has no factory floor and no robot on the market under its own name. Its entire future depends on robot manufacturers choosing to run Pi’s brain inside their machines rather than building their own, and several of the largest humanoid robot companies in the world, including Figure AI, are already doing exactly that. Competition from open-weight alternatives is intensifying, and the leading vendor list in this category has shifted at least once every month through early 2026. Pi is building something the industry has never seen before, and whether one company gets to own that layer permanently is a question the quality of the work alone can not settle.

About The Author

I help AI companies communicate what their products actually do, clearly, honestly, and in language that respects the reader’s intelligence. If your product is more capable than the way it is currently being described, let us fix that together.

Continue Reading
→ From Watching to Knowing: A Robot’s Quiet Mastery of Its Hands
→ The Cost of AI Over-Reliance Is Paid in Human Thinking

Favour Ogoba · @favourogoba67 on Medium · favourogoba67-crypto.github.io

A factory floor robot can pick up a sheet metal panel a thousand times a day without error, and then fail completely the moment someone hands it a stapler. Nothing about the stapler is mechanically harder to grip than the panel. The robot simply was never taught it. Every motion an industrial robot performs has to be coded by an engineer in advance, one task, one object, one exact set of coordinates at a time. Change the object and the robot becomes useless, no matter how strong or precise its hardware happens to be. This is the wall most of robotics has been building against for years. Figure AI built something designed to walk straight through it.

A Stapler Breaks the Old Way of Teaching Robots

Figure AI is the company behind the humanoid robots increasingly seen in factory videos and conference demos, but the robot itself is not really the product. The body, called Figure 03, is a refined piece of hardware with a faster camera system and a wider field of view than its predecessor. What makes the company’s machines genuinely useful is something most coverage skips past entirely: the model running inside them, called Helix.

Figure Sells a Mind Not Just a Machine

Helix is what is known as a vision-language-action model, a single neural network trained to do three things at once: see what is in front of it, understand a sentence spoken in plain English, and convert that understanding directly into movement. It is the first model of its kind to control a humanoid’s entire upper body in one motion, including the wrists, torso, head, and individual fingers, and it is also the first capable of running on two robots simultaneously so they can complete one shared task using objects neither has encountered before. A robot that can only repeat what it has already shown stays a tool. Helix is built to act on what it has never seen, which is what starts to make it feel like a coworker.

Spartanburg Proved the Concept Under Real Pressure

The clearest proof that this holds up outside a lab sits on BMW’s assembly line in Spartanburg, South Carolina. Two Figure 02 robots ran ten-hour shifts five days a week for eleven months, accumulating 1,250 hours of runtime while loading more than 90,000 sheet metal parts and contributing to the production of more than 30,000 BMW X3 vehicles. The task sounds modest on paper: pick a metal part off a rack and place it precisely onto a welding fixture. In practice, the robots were held to a 37-second load time inside an 84-second total cycle, a pace that left almost no room for hesitation or error.

The pilot did not pretend to be flawless. Figure later confirmed that the robot’s forearm was the single most common point of hardware failure during the deployment, a finding the company says was fed directly into redesigning the wrist electronics on Figure 03. An eleven-month run on a real production line, with real failure data and a real fix that followed, carries more weight than any demo reel.

One System Thinks While the Other Moves

The mechanism behind this is the one technical detail worth knowing, because it explains why Helix feels less like a script and more like instinct. Helix runs two systems together: a larger model with roughly 7 billion parameters handles scene understanding and language at a slower 7 to 9 times per second, while a much smaller model, around 80 million parameters, handles the actual physical movement at 200 times per second. The slow half decides what should happen, and the fast half carries it out before that decision even has time to settle.

The Body Moves Before the Mind Finishes

Picture a chef glancing at a recipe, deciding what to cook, and then chopping vegetables without consciously thinking through each cut. The decision and the motion run on two different clocks inside the same person. Helix is built the same way. One part of Helix reads the room and the request, and almost automatically, the rest of it already knows how a hand should close around an object it has only just seen.

Even Its Builder Is Not Fully Convinced Yet

What keeps this story honest is that the people building it are not claiming the job is done. Figure AI’s founder, Brett Adcock, has had robots running in his own home around his children for over a year, closely supervised the entire time. Asked directlywhether he would let one operate freely near his kids without anyone watching. He said no, not yet. The robots that have learned to reach toward the unfamiliar still have not earned the kind of trust that comes only with time, and the company building them seems to know it.

What makes Helix worth knowing has nothing to do with solving physical AI. For the first time, watching led somewhere a robot’s body could actually follow.

About The Author

I help AI startups communicate what their products really do — clearly, honestly, and in language, their users actually understand. If your product is more powerful than the way it is currently being described, let us fix that together.

Continue Reading
→ How Read AI Built Ada to Work on Your Behalf When You Cannot Be There
→ What Perplexity Computer Gets Right About How Professionals Actually Search
Favour Ogoba · @favourogoba67 on Medium · favourogoba67-crypto.github.io

From Watching to Knowing: A Robot’s Quiet Mastery of Its Hands was originally published in Reflections & Realities on Medium, where people are continuing the conversation by highlighting and responding to this story.

This is a plain-language rewrite of the original Jasper Agents and Canvas announcement. Read the original here: Jasper Ushers in the Agentic Era of Marketing, PRNewswire (June 2025).

Ask a marketing manager what Jasper does and they will tell you it writes content. Ask the same question of someone at Prudential, Wayfair, or Ulta Beauty and the answer sounds different. Those companies are not using Jasper to write faster. They are using it to run the entire content operation, from brief to approval to publish, across every channel, every region, and every campaign simultaneously. A different product category is wearing the same familiar name.

Jasper made that shift official at Cannes in June 2025, launching Agents, Canvas, and a rebrand that declared it the world’s leading agentic marketing platform. The writing tool that helped a marketing manager defeat the blank page now runs the operation that surrounds every writer on the floor.

One Pair of Hands Got Faster: The Room Did Not

Every AI generation tool that landed on a marketing team’s desk between 2021 and 2024 solved the same problem, the opening sentence that faced the writer every morning. A writer opened the software, described what they needed, and received a draft in seconds while output accelerated beyond anything the team had planned for. The individual became a one-person content engine capable of producing in a day what previously demanded a full week.

The pipeline refused to move at the same pace. Briefs still waited for creative directors who were already stretched beyond capacity. Drafts still sat in approval queues moving at the pace of calendar gaps. Published copy still drifted from brand guidelines because nobody carried the bandwidth to enforce them consistently across every piece, every channel, and every market. Faster generation filled the queue with greater urgency and the queue absorbed it all without shortening.

Cannes Was Where Jasper Stopped Apologising for Its Ambition

Jasper was founded in January 2021 by Dave Rogenmoser, Chris Hull, and John Morgan, three friends who had spent eight years building companies together before finding the product that fit. Their first agency client paid $300 a month for SEO work on a porta-potty rental business. Their AI writing tool reached $42.5 million in annual recurring revenue within its first year.

The company changed names twice, navigated a leadership transition in 2023, and arrived at Cannes in June 2025, carrying something its earlier versions were never architected to be. CEO Timothy Young described the moment as the opening of the agentic era of marketing. What walked into that room was a writing instrument. What walked out was a platform designed to govern the entire content lifecycle from a single connected surface.

The Bigger the Brand, the Louder the Queue

The challenge Jasper addresses expands with the size of the organisation carrying it. A three-person startup produces copy as the moment demands it. A global company with regional marketing teams, compliance requirements, multi-language campaigns, and dozens of simultaneous product launches generates material at a volume no individual tool can coordinate. The brief enters the room and every agent, writer, reviewer, and approver needs to touch it before it reaches the audience. Every handoff adds weight to a process already carrying too much. Every added weight is a campaign that ships late or arrives off-brand.

Jasper is trusted by nearly 20% of the Fortune 500, including Prudential, Ulta Beauty, and Wayfair. These organisations moved beyond AI writing experiments into something more structural, running output pipelines at a scale where the coordination gap is larger than the creation gap. Jasper rebuilt itself to close that specific distance.

The creation gap was already closing across the industry. The coordination gap was the one nobody had touched.

The Operation Has Three Chambers and Nothing Falls Between Them

Jasper’s June 2025 launch introduced three connected capabilities designed to run the full content lifecycle without a single gap between them. Jasper Agents are purpose-built marketing agents that handle specific tasks autonomously, research, briefing, drafting, and optimisation, operating with full knowledge of the brand’s voice, guidelines, and audience so that every output arrives already shaped for the company it represents.

Jasper Canvas is the workspace where human and agent work converges. Briefs, drafts, feedback, and approvals all occupy the same surface, ending the scatter of email threads and shared documents that fragments most content workflows. Jasper Grid, launched in November 2025, scales the system into a structured sheet-like interface that allows teams to create and orchestrate work across every channel, audience, and region simultaneously. The brief walks in through one end of the operation. The campaign walks out through the other end fully formed. The handoffs that once consumed days now consume minutes.

The Writer Accelerated Into a Wall That Had Always Been There

The analogy that makes this shift visible is the difference between giving a chef a sharper knife and redesigning the kitchen. A sharper knife helps the person holding it. A redesigned kitchen changes what the entire team can produce in a single service. Jasper before 2025 was the sharper knife, delivering faster output for the individual reaching for it. Jasper after June 2025 is the redesigned kitchen, a coordinated environment where every role in the content workflow has a designated station and a unified path forward.

Speed without structure deposits faster chaos into a system already at capacity. The marketing teams that absorbed AI writing capabilities without addressing the structural challenge underneath found their queues growing heavier as their individual writers grew more prolific.

The bottleneck climbed upstream every time the creation dimension accelerated without a corresponding shift in the coordination dimension. Jasper’s move from writing instrument to operational infrastructure is a response to both dimensions simultaneously.

A sharper knife improves the person holding it. A redesigned kitchen transforms what the whole team can deliver.

What the Team Became Did Not Wait for Permission From the Org Chart

Jasper’s 2026 State of AI in Marketing Report surveyed 1,400 marketers and found that only 41% could confidently prove AI ROI, down from 49% the previous year, even as adoption had become nearly universal. The instruments were the evidence that their value was not keeping pace with their spread.

The marketing teams moving beyond individual AI capabilities and into operational platforms are the ones beginning to close that divide. When the brief, the draft, the review, the approval, and the publish all run through a unified workspace with a shared understanding of the brand, the measurement question becomes answerable because the process is no longer fragmented across a dozen competing tools.

Jasper spent four years as the capability that made individual writers more prolific. The platform it became in June 2025 was designed for teams that understood the real constraint was never the writer. The constraint was the distance between the brief and the audience, and every unnecessary handoff that lived in between. The team that closes that distance before its competitors do will not need to wait for the org chart to catch up.

About The Author

I help AI startups communicate what their products really do — clearly, honestly, and in language their users actually understand. If your product is more powerful than the way it is currently being described, let us fix that together.

Continue Reading
→ Cohere Hit $240 Million Without Reading a Single Customer File
→ The Cost of AI Over-Reliance Is Paid in Human Thinking

Favour Ogoba · @favourogoba67 on Medium · favourogoba67-crypto.github.io

The AI Writing Tool That Became a Marketing Department was originally published in Reflections & Realities on Medium, where people are continuing the conversation by highlighting and responding to this story.

This is a plain-language rewrite of the official V7 release documentation. Read the original here: Midjourney Version Documentation, Midjourney (April 2025).

Type a prompt into Midjourney and it generates an image. Type the same prompt into a colleague's account and the output looks completely different. Same words, same tool, and yet the results look like they came from two different people with two different eyes. V7 is the version that explains why they did.

On April 3, 2025, Midjourney released V7, the first version of itself that learns what each user finds visually beautiful and applies that understanding automatically. The prompt describes the subject, and the model already knows how you want it to feel.

Every Midjourney Prompt Before V7 Was a Shot in the Dark

Writing the right prompt was the entire job. The more precisely you described the lighting, the mood, the composition, the colour palette, the artistic reference, the camera angle, and the texture, the closer the output came to what you had in mind. Skilled Midjourney users spent years developing prompt literacy, learning the specific vocabulary the model responded to, the modifiers that moved output in predictable directions, the combinations that produced reliable outputs.

The tool held no memory of what you had generated before, no record of what you kept and what you discarded, no understanding of whether you preferred warm tones over cool ones, graphic compositions over painterly ones, or sparse scenes over dense ones. Every generation was addressed to a stranger who had never seen your work.

David Holz Called V7 a Different Architecture and Meant It

Holz confirmed in the official release that the team rebuilt V7 from scratch with a new underlying design, producing improvements that were immediate and measurable across every output category. Text and image prompts are handled with greater accuracy. Image quality delivers richer textures and more coherent details across complex elements, including bodies, hands, and objects, historically the most unreliable part of AI image generation. V7 also introduced Draft Mode for rapid iteration and Omni Reference for consistent character and object rendering across multiple generations.

The rebuild from scratch made both the performance gains and the personalisation system achievable, something the previous version could not have supported at scale. Holz rebuilt V7 to create the conditions under which the tool could learn, and the quality improvements arrived as a consequence of that decision.

The First Model That Learns What You Find Beautiful by Default

V7 introduced personalisation by default. The first time, any Midjourney model activated this feature automatically and without requiring users to opt in. To build a personalisation profile, users rate approximately 200 images from Midjourney’s library, selecting which they find more visually appealing across each comparison. The ratings require no explanation or categorisation from the user. The system observes the pattern across the full set and constructs a model of each user’s visual preferences.

Once the profile exists, every generation applies it silently. The model interprets each new prompt through the lens of what it has learned about that specific user’s eye. Two users writing identical prompts will receive different outputs because the tool now holds a different model of each person’s visual sensibility. The prompt entered is identical in both cases, and the people behind it carry completely different visual histories.

Two Hundred Images Later the Tool Stops Guessing About You

Completing the rating process takes approximately fifteen to twenty minutes. Users see pairs of images and select which they find more appealing across roughly 200 comparisons. The selections require no explanation or justification from the user. The pattern of choices across that many comparisons produces a reliable signal about the user’s preferences, whether they gravitate toward painterly or photographic aesthetics, high contrast or muted palettes, figurative or abstract compositions, dense visual information or open space.

The profile updates as the user continues generating and rating. Preferences shift over time, and the model adjusts with them. A user whose taste evolves toward a different aesthetic over six months will find their generations shifting to reflect that evolution without reconfiguring any settings. The tool learns continuously from everything the user produces and keeps.

Two users. Identical prompts. The outputs look like they came from two different people because the tool knows they are.

A Stylist Who Has Dressed You a Thousand Times Already Knows

Think about the difference between walking into a clothing shop staffed by someone who has never met you and walking into one staffed by someone who has dressed you for a decade. The stranger asks what you are looking for and produces options based on your answer. The person who knows you pulls three things off the rail before you finish the sentence, because the pattern of what you have chosen over years has given them a model of your taste that no single description could replicate.

Midjourney V7 is the person who has dressed you a thousand times. The 200-image rating session is the first fitting. Every generation after that draws from a growing record of what you have kept, what you have discarded, and what the pattern across both reveals about how you see. The prompt carries what you want to show. The personalisation profile shapes how you want it to feel. Those are two different kinds of instruction and V7 is the first model equipped to handle both simultaneously.

The Prompt Gets Shorter When the Tool Already Knows Your Eye

V7 became the default Midjourney model on June 17, 2025. V8 Alpha launched on March 17, 2026, on alpha.midjourney.com, with V8.1 following on April 30, 2026, as the fastest model in Midjourney’s history. Both build on the personalisation foundation V7 established.

Personalisation has moved from an optional feature to the starting point every generation now runs from.
The practical consequence is a different kind of creative workflow. A user with a developed personalisation profile writes shorter prompts and produces more accurate outputs, because the tool fills in the unstated preferences from its model of their taste. The prompt carries the subject and the tool applies the aesthetic, compressing a collaboration that once required hundreds of words of description into tens.

Midjourney spent years teaching users to write better prompts. V7 reframed what the tool needed from the person using it. Articulating your own aesthetic precisely enough for a blank model to understand is a skill that takes time to develop. V7 removed that requirement entirely from the workflow. The tool learned your aesthetic from watching what you chose.

About The Author

I help AI startups communicate what their products really do — clearly, honestly, and in language, their users actually understand. If your product is more powerful than the way it is currently being described, let us fix that together.

Continue Reading

→ Runway Makes AI Video: So Why Is It Trying to Cure Disease
→ The Video That Talks Back: What Synthesia 3.0 Actually Built

Favour Ogoba · @favourogoba67 on Medium · favourogoba67-crypto.github.io

This is a plain-language rewrite of the original GWM-1 announcement. Read the original here: Introducing Runway GWM-1, Runway Research (December 2025).

Runway is the company that gave filmmakers AI-generated video. Hollywood uses it. Advertising agencies use it. Students use it to produce work that previously required a production team. If you had to describe Runway in one sentence, you would say it makes video.

On December 11, 2025, Runway announced GWM-1, its first General World Model. The announcement described ambitions that have nothing to do with video production. Robotics. Disease. Scientific discovery. A company whose product sits in creative workflows quietly declared that the next frontier of artificial intelligence requires something language models can not deliver. That declaration is worth understanding.

ChatGPT Can Describe a Bicycle but Cannot Ride One

Ask a language model to describe how a bicycle stays upright and it will give you a technically accurate answer about angular momentum, gyroscopic forces, and the geometry of the steering axis. The description will be correct. The model has no idea what any of it feels like.

This is the ceiling that Runway named directly in the GWM-1 announcement: language models alone will not solve the world’s hardest problems. Robotics requires a machine that understands what happens when a wheel loses traction. Disease research requires a model that can simulate how a protein folds under different conditions. Scientific discovery requires systems that can run experiments in a virtual environment and learn from what goes wrong. Describing those processes in text produces nothing close to experiencing them. The next generation of AI needs to have lived inside physical reality, even if that reality is simulated.

Runway Turned Its Video Model Into a Reality Engine

Runway has been generating video since 2023. Its Gen-4.5 model, released in December 2025, holds the number one position on the Video Arena leaderboard maintained by Artificial Analysis, surpassing models from Google and OpenAI. CEO Cristóbal Valenzuela told CNBC: "We managed to out-compete trillion-dollar companies with a team of 100 people."

GWM-1 is built on top of Gen-4.5. The model that generates cinematic video became the foundation for a system that simulates physical reality. Runway did not abandon video. It discovered that a model trained to understand how objects move, how light behaves, and how physics governs motion in a frame was already most of the way toward a model that could simulate a world. The video model was the prototype for the reality engine.

Three Variants, Three Problems Language Models Left Unsolved

GWM-1 ships in three variants, each addressing a category where language models reach their limit. GWM Worlds generates real-time explorable environments from a prompt or image. Set a scene and move through it — the model generates geometry, lighting, and physics as you go, maintaining spatial consistency across the full sequence. The environment reacts to defined rules. Prompt flight, and the agent leaves the ground. Prompt gravity, and it stays on it.

GWM Robotics generates synthetic training data for robot policies, allowing engineers to test how a robotic system behaves across thousands of scenarios without deploying physical hardware. Runway Characters, released in May 2026, is an audio-driven interactive model that generates realistic conversational characters from a single image. The character listens, responds, maintains eye contact, and holds extended conversations without quality degradation. All three variants run on the same underlying world model trained to understand how reality behaves.

Three variants. Three domains where describing a problem was never going to be enough to solve it.

GWM-1 Learns Physics the Way a Child Learns to Fall

GWM-1 is an autoregressive model. It generates the world one frame at a time, predicting each new frame based on everything that came before it. The prediction is conditioned on actions: camera movement, robot commands, speech input , which means the model generates a world that responds to what happens inside it.

The physics GWM-1 understands was learned by watching video at scale. Millions of hours of footage showing how water flows, how objects fall, how surfaces behave under pressure, how humans move — all of it absorbed through the same training process that taught Gen-4.5 to generate realistic video. The model did not study physics. It watched the world until physics became something it could predict.

The Difference Between Reading About Fire and Getting Burned

Think about how a child learns that fire is dangerous. A textbook describes combustion, temperature thresholds, and tissue damage. The child reads and understands the words. The understanding stays abstract until the child gets too close to a candle and feels the heat on their skin. That moment of physical experience encodes the lesson in a way no description can replicate.

Language models have read everything ever written about fire. They can explain combustion with precision. They have never felt heat. GWM-1 was trained in the equivalent of a room full of candles — simulated environments where the consequences of proximity to heat play out frame by frame, where objects behave the way objects behave, where reality is generated rather than described. The difference between those two kinds of knowledge is exactly the gap Runway is trying to close.

A language model knows everything written about the physical world. GWM-1 has lived inside a simulation of it.

The Hardest Problems Are Waiting for AI That Has Felt the World

Runway stated its position plainly in the GWM-1 announcement: "Language models alone won’t solve the world’s hardest problems — robotics, disease, scientific discovery. Real progress requires models that experience the world and learn from their mistakes, the same way that humans do."

A robotics company using GWM Robotics can run a robot policy through ten thousand simulated environments before physical hardware is ever switched on. A research team can generate synthetic training scenarios that would take years and millions of dollars to collect in the real world. A product team can build a conversational character that holds genuine real-time dialogue rather than retrieving scripted responses.
Runway is a video company that spent seven years learning how reality moves through a frame. GWM-1 is what that knowledge becomes when the goal shifts from showing the world to simulating it. The problems waiting on the other side of that shift are not creative ones.

About The Author

I help AI startups communicate what their products really do — clearly, honestly, and in language their users actually understand. If your product is more powerful than the way it is currently being described, let us fix that together.

Continue Reading
→ The Video That Talks Back: What Synthesia 3.0 Actually Built
→ The Cost of AI Over-Reliance Is Paid in Human Thinking

Favour Ogoba · @favourogoba67 on Medium · favourogoba67-crypto.github.io