Why Continual Learning Keeps AGI Further Off Than It Looks: Dwarkesh's Timelines as of July 2025
Dwarkesh Patel
On the Dwarkesh Podcast, guests have given very different estimates of how far away AGI is. Some say 20 years, others say two. In this solo episode, adapted from a blog post, Dwarkesh sets out their own view as of July 2025. They argue that today's large language models are impressive but lack a basic capability, the ability to learn on the job. They think this makes near-term economic transformation less likely than some researchers expect. They also think that once the problem is solved, the change could be sudden and enormous.
Why today's models aren't already transforming the economy
Dwarkesh starts by rejecting a common claim: that even if all AI progress stopped today, current systems would still change the economy more than the internet did. They call current LLMs "magical." But they do not think the Fortune 500 is slow to reorganize around them because management is too conservative. Their view is that getting ordinary, humanlike work out of these models is genuinely hard, because the models lack some fundamental capabilities.
This view comes partly from direct experience. Dwarkesh calls themselves "AI forward" and estimates they have spent more than a hundred hours building small LLM tools for the podcast's post-production. They have asked models to:
- rewrite auto-generated transcripts so they read the way a human editor would make them read,
- pick out clips from a transcript,
- co-write essays with them passage by passage.
They describe these as simple, self-contained, short-horizon, language-in/language-out tasks that should sit right at the center of what LLMs do well. They rate the models about 5 out of 10 on them. They say that is impressive. They also say the experience of trying to make these tools useful has lengthened their timelines.
The core problem: models don't get better over time
Dwarkesh names the lack of continual learning as "a huge huge bottleneck." On many tasks an LLM's starting level may be higher than an average person's. But there is no way to give the model high-level feedback and have it improve. You are stuck with whatever it can do out of the box. You can keep adjusting the system prompt, but in practice that produces nothing close to the learning and improvement human employees go through.
In Dwarkesh's view, humans are useful mainly because of this learning, not because of raw intelligence. People build up context, examine their own failures, and pick up small improvements and efficiencies as they repeat a task.
They illustrate this with an analogy about learning the saxophone. Normally a child blows into the instrument, hears the result, and adjusts. Now imagine a different method. A student makes one attempt. At the first mistake, the student is sent away and the teacher writes detailed notes on what went wrong. The next student reads the notes and tries to play Charlie Parker cold. When that student fails, the teacher refines the notes and calls in another. Dwarkesh says this would never work, however well the instructions are written. Yet it is essentially the only way we currently have to "teach" an LLM anything.
They acknowledge that RL fine-tuning exists. They argue it is not a deliberate, adaptive process the way human learning is. They point to their own editors, who have become extremely good. The editors did not get there because someone built custom RL environments for each subtask of their job. They noticed many small things themselves and thought hard about what resonates with the audience, what content Dwarkesh likes, and how to improve their daily workflow.
Could a smarter model train itself?
Dwarkesh considers one way out. A smarter model might build its own RL loop in a way that looks organic from the outside. The user gives high-level feedback, and the model invents verifiable practice problems to train on, or even builds a whole environment to rehearse the skills it thinks it lacks. Dwarkesh says this "just sounds really hard." They are unsure how well such techniques would generalize across different kinds of tasks and feedback.
They expect models to eventually learn on the job the way humans do. They find it hard to see that happening in the next few years, because they see no obvious way to add continual learning to the kind of models LLMs are.
Learning within a session, and losing it
Dwarkesh notes that LLMs do become fairly smart and useful within a single session. When co-writing an essay, they give the model an outline and ask for drafts one passage at a time. Up to about the fourth paragraph, the suggestions are all bad. Dwarkesh rewrites each paragraph from scratch and tells the model, in their words, "Look, your shit sucked. This is what I wrote instead." After that, the model starts giving good suggestions for the next paragraph. But this subtle grasp of their preferences and style is gone once the session ends.
One possible fix is a long, rolling context window, like the one Claude Code already has, which compacts session memory into a summary every 30 minutes. Dwarkesh expects that turning rich tacit experience into a text summary will be brittle outside software engineering. Software is heavily text-based, and the codebase itself already serves as an external memory. They return to the saxophone: imagine teaching a child to play from text alone. They add that even Claude Code often undoes a hard-won optimization the two of them built together before a /compact, because the reason for the change didn't make it into the summary.
Disagreeing with Sholto Douglas and Trenton Bricken
This reasoning is why Dwarkesh disagrees with a claim that Anthropic researchers Sholto Douglas and Trenton Bricken made on the podcast. Dwarkesh quotes Trenton: "Even if AI progress totally stalls (and you think that the models are really spiky, and they don't have general intelligence), it's so economically valuable, and sufficiently easy to collect data on all of these different white collar job tasks, such that to Sholto's point we should expect to see them automated within the next five years."
Dwarkesh's own estimate is that if AI progress stopped today, less than 25% of white-collar employment would disappear. Many tasks would be automated. For example, Claude 4 Opus can technically rewrite auto-generated transcripts for them. But because it cannot improve over time or learn their preferences, they still hire a human for the job. They expect the same pattern across other white-collar work, even with more data, unless continual learning improves. AIs will be able to do many subtasks somewhat satisfactorily. But because they cannot build up context, Dwarkesh argues, they will not be able to work as real employees inside a firm.
Bearish now, very bullish later
Dwarkesh says the same reasoning that makes them bearish on transformative AI in the next few years makes them especially bullish over the coming decades. When continual learning is solved, they expect a large jump in how valuable these models are.
They say this could happen even without a "software-only singularity," where models rapidly build ever-smarter successors. The result might instead look like a broadly deployed intelligence explosion. AIs would work throughout the economy, doing different jobs and learning as they go the way humans do. Unlike humans, the copies could combine what they learn, so in effect one AI would be learning every job in the economy. Dwarkesh suggests that an AI capable of this kind of online learning might quickly become superintelligent even with no further algorithmic progress.
They do not expect this to arrive all at once, for example through an OpenAI livestream announcing that continual learning is solved. Labs have strong incentives to release innovations quickly. So Dwarkesh expects to see a broken early version of continual learning, or "test time training, or whatever you want to call it," before anything that truly learns like a human. They expect plenty of warning before this bottleneck is fully removed.
Skepticism about reliable computer-use agents by next year
Sholto and Trenton also told Dwarkesh they expect reliable computer-use agents by the end of the following year. Computer-use agents already exist, but Dwarkesh calls them "pretty bad." The researchers had something far more capable in mind. You would tell an AI, "Go do my taxes," and it would go through your email, Amazon orders, and Slack messages. It would email everyone you need invoices from, compile your receipts, decide which items are business expenses, ask for your approval on edge cases, and then submit Form 1040 to the IRS.
Dwarkesh is skeptical. They stress that they are not an AI researcher and do not want to contradict researchers on technical details. With that caveat, they give three reasons they would bet against the forecast.
First, longer horizons mean longer rollouts. The AI may need to do two hours of agentic computer work before anyone can tell whether it succeeded. Computer use also involves processing images and video, which already costs more compute, before counting the longer rollouts. Dwarkesh thinks this should slow progress.
Second, there is no large pretraining corpus of multimodal computer-use data. They quote a post from Mechanize on automating software engineering: "For the past decade of scaling, we've been spoiled by the enormous amount of internet data that was freely available for us to use. This was enough to crack natural language processing, but not for getting models to become reliable, competent agents. Imagine trying to train GPT-4 on all the text data available in 1980—the data would have been nowhere near enough, even if you had the necessary compute." Dwarkesh names possibilities that would weaken this argument. Text-only training might already give models a strong sense of how UIs work and how their parts relate. RL fine-tuning might be sample-efficient enough that little data is needed. Models might be good enough at front-end coding to generate millions of toy UIs to practice on. But Dwarkesh says they have seen no public evidence that models have suddenly become less data-hungry, especially in domains where they have had much less practice.
Third, even ideas that look simple in hindsight take a long time to get working. The RL procedure DeepSeek described in its R1 paper looks simple at a high level. Yet it took two years from GPT-4's development and launch to the release of o1. Dwarkesh admits it would be "insanely and hilariously arrogant" to call R1 or o1 easy. Reaching them took a great deal of engineering, debugging, and ruling out alternative ideas. They say that is exactly their point. If it took that long to implement "train a model to solve verifiable math and coding problems," we are probably underestimating how hard computer use will be, since it involves a different modality and much less data.
The progress that is real
Dwarkesh then deliberately turns away from the skepticism. They say they don't want to be like the "spoiled children on Hackernews" who, handed a goose that lays golden eggs, would complain about how loud it quacks.
They point to the reasoning traces of o3 and Gemini 2.5 and say the models really are reasoning. They break down problems, think about what the user wants, respond to their own internal monologue, and correct themselves when they notice a line of thought is unproductive. Dwarkesh finds it strange that people treat this as ordinary, as if machines naturally go off, think, and return with a smart answer.
They suggest some people are too pessimistic because they haven't used the strongest models in the areas where those models are best. As an example, they describe giving Claude Code a vague spec and waiting ten minutes while it builds a working application on the first try, which they call "a wild experience." You could explain this in terms of circuits, training distributions, or RL. Dwarkesh says the most direct, concise, and accurate explanation is that it is "powered by a baby general intelligence." At that point, they say, part of you has to think: "It's actually working. We're making machines that are intelligent."
The predictions: 2028 and 2032
Dwarkesh says their probability distributions are very wide and that they take probability distributions seriously. So they think preparing for a misaligned superintelligence in 2028 still makes a lot of sense, and they consider that a fully plausible outcome. They then give the years at which they would take a 50/50 bet.
Computer use: 2028. The milestone is an AI that can do their small business's taxes end to end, as well as a competent general manager could in a week. That includes tracking down receipts on various websites, finding missing pieces, emailing people for invoices, filling out the forms, and filing with the IRS. Dwarkesh thinks computer use is currently in its "GPT-2 era." There is no pretraining corpus, and models must optimize for a much sparser reward over a much longer horizon, using action primitives they are not familiar with. On the other side, the base model is already fairly smart and may have a useful prior for computer-use tasks. There is also far more compute and many more AI researchers than before, so these factors might balance out. They see small-business tax preparation as the computer-use equivalent of what GPT-4 was for language, and it took four years to get from GPT-2 to GPT-4. They clarify that they expect impressive computer-use demos in 2026 and 2027. GPT-3 was very impressive but not very useful in practice. Their claim is only that models before 2028 won't be able to handle, end to end, a week-long, fairly involved project that requires computer use.
On-the-job learning: 2032. This milestone is an AI that can learn on the job as easily, organically, seamlessly, and quickly as a human, in any white-collar role. Their example is an AI video editor that, after six months, understands their preferences, the channel, and what works for the audience as deeply and usefully as a human editor would. They still see no obvious way to add continuous online learning to LLMs. But they note that seven years is a long time: GPT-1 had just come out seven years earlier. It does not seem implausible to them that some method will be found in that period.
Why timelines are "this decade or bust"
Dwarkesh anticipates an objection. They have argued that the lack of continual learning is a major handicap, yet they are predicting that in seven years we could see something that at minimum looks like a broadly deployed intelligence explosion. They accept this. They say they are forecasting a very strange world within a fairly short time.
Their reason is that AGI timelines are "very lognormal." They sum it up as "It's either this decade or bust," then qualify it: the more accurate claim is that the probability per year falls over time, which is "less catchy." Over the past decade, AI progress has come from scaling training compute on frontier systems by more than 4x per year. Dwarkesh argues this cannot continue past this decade, whether you look at chips, power, or the share of GDP spent on training. After 2030, progress will have to come mostly from algorithmic advances, and even there the easy gains will be used up, at least within the deep learning paradigm. As a result, the yearly probability of AGI drops sharply.
Dwarkesh draws two conclusions from this. If the longer side of their 50/50 bets turns out to be right, the world could stay relatively normal into the 2030s or even the 2040s. In all the other scenarios, even if we stay clear-eyed about AI's current limits, they say we should expect some truly crazy outcomes.
Dwarkesh ends by explaining where the essay came from. It began as a blog post after their conversation with Sholto and Trenton. During the interview they found they disagreed on timelines, and it took several weeks of reflection afterward to work out exactly where they disagreed and why their own timelines were longer.
I’ve had a lot of discussions on my podcast where we haggle out our timelines to AGI. Some guests think it’s 20 years away. Others 2 years. Here’s where my thoughts lie as of July 2025.
Sometimes people say that even if all AI progress totally stopped, the systems of today would still be far more economically transformative than the internet. I disagree. I think the LLMs of today are magical. But the reason that the Fortune 500 aren’t using them to totally transform their workflows isn’t because the management there is too stodgy. Rather, I think it’s genuinely hard to get normal humanlike labor out of these LLMs. And this has to do with some fundamental capabilities these models lack.
I like to think I’m "AI forward" here at the Dwarkesh Podcast. I’ve probably spent over a hundred hours trying to build these little LLM tools for my post production setup. And the experience of trying to get them to be useful has extended my timelines.
I’ll try to get LLMs to rewrite autogenerated transcripts for me, to optimize for readability in the way a human would. Or I’ll try to get them to identify clips from a transcript that I feed in. Sometimes I’ll try to get them to co-write an essay with me, passage by passage. These are simple, self contained, short horizon, language in, language out tasks—the kinds of assignments that should be dead center in the LLMs’ repertoire. And they're 5/10 at them. Don’t get me wrong, that is impressive.
But the fundamental problem is that LLMs don’t get better over time the way a human would. This lack of continual learning is a huge huge bottleneck. The LLM baseline at many tasks might be higher than the average human's. But there’s no way to give a model high level feedback. You’re stuck with the abilities you get out of the box. You can keep messing around with the system prompt, but in practice this just doesn’t produce anything even close to the kind of learning and improvement that human employees experience.
The reason humans are so useful is not mainly their raw intellect. It’s their ability to build up context, interrogate their own failures, and pick up small improvements and efficiencies as they practice a task.
How would you teach a kid to play the saxophone? You'd have them try to blow into one, and then they'd see how it sounds, and they'd adjust. Now imagine if this was the way you'd have to teach saxophone instead: A student takes one attempt. And the moment they make a mistake, you send them away and you write detailed instructions about what went wrong. And you call the next student in. And the next student reads your notes and tries to play Charlie Parker cold. And when they fail, you refine your instructions and you invite the next student. This just wouldn’t work. No matter how well honed your prompt is, no kid is just going to learn how to play saxophone from reading your instructions.
But this is the only modality we have to ‘teach’ LLMs anything. Yes, there’s RL fine tuning. But it’s not a deliberate, adaptive process in the way that human learning is.
My editors have gotten extremely good. And they wouldn’t have gotten that way if we had to build bespoke RL environments for different subtasks involved in their work. They’ve just noticed a lot of small things themselves and thought hard about what resonates with the audience, what kind of content I like, and how they can improve their day to day workflows.
Now, it’s possible to imagine ways in which a smarter model could build a dedicated RL loop for itself which just feels super organic from the outside. I give some high level feedback, and the model comes up with a bunch of verifiable practice problems to RL on—maybe even a whole environment in which it gets to rehearse the skills that it thinks it's lacking. But this just sounds really hard. And I don’t know how well these techniques will generalize to different kinds of tasks and feedback.
Eventually the models will be able to learn on the job in this organic way that humans can. But it’s just hard for me to see how that could happen within the next few years, given there’s no immediately obvious way in which to slot in continuous learning into the kinds of models that these LLMs are.
LLMs actually do get kinda smart and useful in the middle of a session. For example, sometimes I’ll co-write an essay with an LLM. I’ll give it an outline, and I’ll ask it to draft the essay passage by passage. And all its suggestions up till paragraph four will just be bad. I'll just rewrite every single paragraph from scratch and tell it, "Look, your shit sucked. This is what I wrote instead." And at this point, it will actually start giving good suggestions for the next paragraph. But this whole subtle understanding of my preferences and style will just be lost by the end of the session.
Maybe there is an easy solution to this that looks like a long rolling context window, like Claude Code already has, which just compacts the session memory into a summary every 30 minutes. I just think that titrating all this rich tacit experience into a text summary will be brittle in domains outside of software engineering, which is very text-based, in which you already have this external scaffold of memory that is stored in the codebase itself.
Again, think about what it would be like to teach a kid to play the saxophone just from text. Even Claude Code will often reverse a hard-earned optimization that we engineered together before I hit /compact—because the explanation for why it was made didn’t make it into the summary.
This is why I disagree with something that Anthropic researchers Sholto Douglas and Trenton Bricken said on my podcast. And this quote is from Trenton: "Even if AI progress totally stalls (and you think that the models are really spiky, and they don't have general intelligence), it's so economically valuable, and sufficiently easy to collect data on all of these different white collar job tasks, such that to Sholto's point we should expect to see them automated within the next five years."
If AI progress totally stops today, I think less than 25% of white collar employment goes away. Sure, many tasks will get automated. Claude 4 Opus can technically rewrite autogenerated transcripts for me. But since it’s not possible for me to have it improve over time and learn my preferences, I still hire a human for this.
So even if we get more data, without progress in continual learning, I think that we will be in a substantially similar position with all other kinds of white collar work. Yes, technically AIs will be able to do a lot of subtasks somewhat satisfactorily, but their inability to build up context will make it impossible to have them operate as actual employees at your firm.
While this makes me bearish about transformative AI in the next few years, it makes me especially bullish on AI over the next decades. When we do solve continual learning, we’ll see a huge discontinuity in the value of these models.
Even if there isn’t a software only singularity, where these models rapidly build smarter and smarter successor systems, we might still get something that looks like a broadly deployed intelligence explosion. AIs will be getting broadly deployed through the economy, and doing different jobs and learning while doing them in the way that humans can. However, unlike humans, these models can amalgamate their learnings across all their copies. So one AI is basically learning how to do every single job in the economy. An AI that is capable of this kind of online learning might rapidly become a superintelligence even if there's no further algorithmic progress.
However, I’m not expecting to watch some OpenAI livestream where they announce that continual learning has been totally solved. Because labs are incentivized to release any innovations quickly, we’ll see a broken early version of continual learning (or test time training, or whatever you want to call it) before we see something which truly learns like a human. I expect to get lots of heads up before this big bottleneck is totally solved.
When I interviewed Anthropic researchers Sholto Douglas and Trenton Bricken on my podcast, they said that they expect reliable computer use agents by the end of next year. We already have computer use agents right now, but they’re pretty bad. They’re imagining something quite different. Their forecast is that by the end of next year, you should be able to tell an AI, "Go do my taxes." And it'll go through all your email, your Amazon orders, and Slack messages, and it will email back and forth with every single person you need to get invoices from, it'll compile all your receipts, decide what things are actually business expenses, and it will ask for your approval on all the edge cases, and then will just submit Form 1040 to the IRS.
I’m skeptical. I’m not an AI researcher, so far be it for me to contradict them on technical details. But given what little I know, here’s why I’d bet against this forecast:
One. As horizon lengths increase, rollouts have to become longer. The AI needs to do two hours worth of agentic computer use tasks before we can even see if it did it right. Not to mention that computer use requires processing images and video, which is already more compute intensive, even if you don’t factor in the longer rollouts. This seems like it should slow down progress.
Two. We don’t have a large pretraining corpus of multimodal computer use data. I like this quote from Mechanize’s post on automating software engineering: "For the past decade of scaling, we’ve been spoiled by the enormous amount of internet data that was freely available for us to use. This was enough to crack natural language processing, but not for getting models to become reliable, competent agents. Imagine trying to train GPT-4 on all the text data available in 1980—the data would have been nowhere near enough, even if you had the necessary compute."
Again, I’m not at the labs. Maybe text only training already gives you a great prior over how different UIs work, and what the relationship is between different components. Maybe RL fine tuning is so sample efficient that you don’t need that much data. But I haven’t seen any public evidence which makes me think that these models have suddenly gotten less data hungry, especially in domains where they’re substantially less practiced.
Alternatively, maybe these models are such good front end coders that they can just generate millions of toy UIs for themselves to practice on.
But, three. Even algorithmic innovations which seem quite simple in retrospect took a long time to iron out. The RL procedure which DeepSeek explained in their R1 paper seems simple at a high level. And yet it took 2 years from the development and launch of GPT-4 to the release of o1.
Now of course I know that it's insanely and hilariously arrogant to say that R1/o1 were easy—a ton of engineering, debugging, and pruning of alternative ideas was required to arrive at this solution. But that’s precisely my point! Seeing how long it took to implement the idea of ‘We should train a model to solve verifiable math and coding problems,’ makes me think that we’re underestimating the difficulty of solving the much gnarlier problem of computer use, where you’re operating in a totally different modality with much less data.
Okay, enough cold water. I’m not going to be like one of those spoiled children on Hacker News who could be handed a golden-egg laying goose and would still spend all their time complaining about how loud its quacks are.
Have you read the reasoning traces from o3 or Gemini 2.5? It’s actually reasoning! It’s breaking down a problem, thinking through what the user wants, reacting to its own internal monologue, and correcting itself when it notices that it's pursuing an unproductive direction. How are we just like, "Oh yeah of course a machine is gonna go think a bunch, come up with a bunch of ideas, and come back to me with a smart answer. That’s just what machines do."
Part of the reason some people are too pessimistic is that they haven’t played around with the smartest models in domains where they’re the most competent. Giving Claude Code a vague spec and just sitting around for 10 minutes while it zero shots a working application is a wild experience. How did it do that? You can talk about circuits and the training distribution and RL or whatever, but the most proximal, concise, and accurate explanation is simply that it’s powered by a baby general intelligence. At this point, part of you has to be thinking, "It’s actually working. We’re making machines that are intelligent."
My probability distributions are super wide. And I want to emphasize that I do believe in probability distributions. Which means that work to prepare for a misaligned 2028 ASI still makes a ton of sense. I think this is a totally plausible outcome. But here are the timelines at which I’d take a 50/50 bet.
An AI that can do taxes end-to-end for my small business as well as a competent general manager could in a week: including chasing down all the receipts on different websites, finding the missing pieces, emailing back and forth with anyone who we need to hassle for invoices, filling out the form, and sending it to the IRS. This I'd say 2028.
I think we’re in the GPT 2 era for computer use. But we have no pretraining corpus, and the models are optimizing for a much sparser reward over a much longer time horizon using action primitives they’re unfamiliar with. That being said, the base model is already decently smart and might have a good prior over computer use tasks, plus there’s a lot more compute and AI researchers in the world, so it might even out. Preparing taxes for a small business feels like for computer use what GPT 4 was for language. It took 4 years to get from GPT 2 to GPT 4.
Just to clarify, I am not saying that we won’t have really cool computer use demos in 2026 and 2027. GPT-3 was super cool, but it was not that practically useful. I’m saying that these models won’t be capable of end-to-end handling a week long and quite involved project which involves computer use.
Ok, and as for the forecast of when AI will be able to learn on the job as easily, organically, seamlessly, and quickly as humans, for any white collar work. For example, if I hired an AI video editor, after six months it would have as much actionable, deep understanding of my preferences, our channel, what works for the audience, as well as a human would. I'd say this would come in 2032.
While I don’t see an obvious way to slot in continuous online learning into the kinds of models these LLMs are, 7 years is a really long time! GPT 1 had just come out this time 7 years ago. It doesn’t seem implausible to me that over the next 7 years, we’ll find some way to get these models to actually learn on the job.
At this point you might be reacting, "Wait you made this huge fuss about continual learning being such a huge handicap. But then your prediction is that we’re 7 years away from what, at a minimum, looks like a broadly deployed intelligence explosion." And yeah, you’re right. I am forecasting a pretty wild world within a relatively short amount of time.
AGI timelines are very lognormal. It's either this decade or bust. (Not really, it's more like lower marginal probability per year—but that’s less catchy). AI progress over the last decade has been driven by scaling training compute on the frontier systems. It's been over 4x a year. This cannot continue beyond this decade, whether you look at chips, power, or even the raw fraction of GDP that is used on training. After 2030, AI progress has to mostly come from algorithmic progress. But even there all the low hanging fruit will be plucked, at least under the deep learning paradigm. So the yearly probability of AGI collapses.
This means that if we end up on the longer side of my 50/50 bets, we might be looking at a relatively
normal world up till the 2030s or even the 2040s. But in all the other worlds, even if we stay sober about the current limitations of AI, we have to expect some truly crazy outcomes.
This was originally a blog post that I published on my website at dwarkesh.com. And it was obviously inspired by the discussion I had with Sholto and Trenton on my podcast, where I ended up disagreeing with them about timelines but it took me a few weeks of thinking afterwards, sorting out exactly where I disagree and why I had longer timelines.
And I do this for other episodes as well. I wrote up some thoughts I had about the many thousands of pages that Stephen Kotkin has written about Stalin, obviously which we were not able to exhaustively cover in that one 2-hour interview.
So anyways, if you want to see these additional artifacts and writing that I produce as a result of this podcast and in preparation for episodes, you should subscribe to my blog and newsletter. You can do that at dwarkesh.com. Otherwise I will also see you next week for a full episode with a real guest.
Article published