Extended Development Record — The Gary Marcus Audit

Excellent. Let's get started on Issue 16. And before we launch the builder draft, since we are aiming for epistemic validity, I'm wondering if we should do a deep web search in an attempt to ensure we understand what Gary Marcus's actual views and claims about AI are, and also what his most inflammatory posts/statements/writing are so we have a grounded view of both before we proceed. What do you think?

Excellent. Before we proceed, are we logging this, including our Signal & Noise discussion from earlier today, for our full attribution record? Before we proceed to the brief/thesis, one thing we might want to do after finalizing Issue 16 and before publishing Issue 2 of Intuition is use all research packet on Marcus to see if we should update anything in Issue 2, and especially to make sure our Issue 2 of Intuition isn't misattributing his claims and statements in any way. Finally, we might want to do a final light NIR that runs across both issues to make sure the two issues are aligned. [Public-page note: access setup from this turn is omitted.]

Excellent. What's our next step now for Signal & Noise Issue 16?

Excellent. I agree. Please proceed.

Excellent. Just curious, wondering if “culturally ineffective” is the best label here, since Gary may be culturally ineffective from my perspective, but not others. Just wondering if there is a way to point out that while this is true, “A critic can be culturally ineffective and epistemically useful at the same time,” it's my personal perception that Gary is culturally ineffective as opposed to a fact that is anchored in any way?

Agree. Please proceed.

Excellent. Can you give me the narrative arc outline first?

OK, that's a nice narrative arc, but as a follow up to Intuition #2, it feels too broad and almost tangential. The core question in Issue 2 is "Do I have to trust Gary Marcus" (even though I would really prefer not to)? My intuition says no, but what does a more rigorous audit of his claims and external reality show? And this is the question that I think Issue 16 should be built upon. We need to fully audit, to the extent reasonably possible, the answer to all these questions we raised in Intuition Issue 2. And ideally in the context of a deep dive into the current state of AI. When the government is shutting down Anthropic due to security risks, do Gary's AI critiques still make sense? My guess is, his critiques probably still do make sense - at least on some levels. Some of them may not and we need to be as clear-eyed and substantiate whether we agree with his actual claims or not. Regardless of how well our deep dive audit into Gary Marcus supports his claims, I'm guessing we could try to land the essay not so much on Marcus as a stress test, but on the unreliability of judging the substance of what people are saying based on how they are saying it, especially when they aren't perceived as part of one's tribe, and ideally, how AI tools, when structured appropriately, can help people filter their feelings from substance/reality. Does this make sense? Can we build an article around this? Or am I reaching too far?

Agree with one clarification. This: “The broad version ‘Marcus is just anti-AI’ should be rejected as unfair” is obvious I think. The critique that I think may hold some water is that Marcus may be technically correct *and* culturally irrelevant - kind of like someone screaming about the dangers of automobiles. And if possible, this “technically correct - at least in some ways - but culturally irrelevant” is the idea that we should stress test.

Can you give me an outline of what the narrative arc would look like now?

Good. Please proceed to the next step.

Overall direction is good. How much can we cut without weakening this issue? Also, if possible, if you haven't already done this, we should probably have a good reference for each claim that we attribute to Marcus. I suspect one of the biggest risks of this essay is misrepresenting Marcus in any way.

Excellent. Agree “Marcus treats continued use as moral failure” is the weakest claim. In fact, I don't think I've ever heard him make this claim or even imply this in any way. However, I think he does treat the idea that LLM's will lead to AGI as an intellectual failure, but again, I also feel like he might be preaching to the choir here. Are there any leading AI labs that think LLM's are the sole path to AGI? Also, “Marcus is culturally irrelevant” is what I would like to believe, obviously not a claim that he has made directly or indirectly. I do think it would sharpen the contrast between what I feel about Gary Marcus and how I should treat his claims if we include direct quotes, linked to external references, of his most inflammatory personal attacks and perhaps political rhetoric. This way readers might actually experience some of the negative feelings, or positive feelings depending on their worldview, that his comments can generate, which could contrast nicely with evidence in support of his views.

OK, my thoughts on the current draft are below. To be honest, I’m not sure how much, if any, of this can be integrated into a coherent essay. Let me know what you think. Consider revising “The more serious critique is not that Marcus is wrong because he is irritating. It is that he may be technically right in some places and still culturally ineffective in others.”, to something closer to “The more serious critique is not that Marcus is wrong because he is irritating. It is that he may be technically right and still culturally ineffective.” Line 33 needs clarification. Consider revising “ The narrower claim is about update power among builders, adopters, investors, and AI-adjacent readers already inclined to dismiss him.”, to something closer to “ The narrower claim is about his warnings completely lacking relevance to the vast majority of builders, adopters, investors, and AI-adjacent readers who have long since moved on.” Line 41 - 43 - We need to drop or modify these lines. I don’t have any suspicion to retire here. I don’t think Gary has ever scolded ordinary users for finding the tools useful - at least to the best of my knowledge. Line 51 - consider changing to something closer to “But does the warning matter when everyone seems to have already accepted to the warning?” It’s like yelling at car manufacturers to install seatbelts. True, but I think they’ve already received the message…” Re: Line 55, the technical issues here are fascinating and the crux of this essay and we shouldn’t just glass this over. Gary’s critiques LLMs, but what alternatives does he support? Neuro-symbolic processing, right? But what is this exactly? How is this different from what LLMS actually do? And What is the evidence to support these Gary’s claims? Are current frontier labs pursuing any of Gary’s recommendations or something similar? Or different than what Gary is recommending but robust and architecturally distinct from LLM’s? If yes, what is the architecture exactly? And we can’t just label/provide a name for any techniques beyond LLMs that frontier AI labs are pursuing, we need to try, in simple terms, to describe/specify exactly what these advanced (non LLM) techniques are doing differently and how they are distinct from LLM processing. In simple terms whenever possible, what techniques/architecture are frontier labs employing and pursuing that might be in full alignment with Gary Marcuses claims about the need for Neuro-symbolic AI? Unless we make the advanced (non-LLM) AI architecture explicitly clear, for all readers will know, all of this “reasoning work, reinforcement learning, test-time thinking, multimodal reasoning, interpretability, alignment, red-teaming, agent autonomy, increasingly capable models, cognitive evaluations, multimodal agents, robotics, planning, and tools, world models, and modular architectures” is just “labels” that obscure the underlying architecture that could still be based on LLM inference. And while we are at it, the crux of the disconnect between Gary’s claim that “LLMs are not the path to AGI” and the government limiting access to Mythos level models made by Anthropic who’s CEO, Dario Amodei, leans hard on scaling laws for increasingly general cognitive capabilities. So there seems there is, at least on the surface, a major disconnect between Gary’s claims that LLMs are unreliable and not a path to AGI *and* the government simultaneously limiting access to a SOTA model because it is so advanced that it presents a major security threat when this SOTA model is made by Anthropic who’s CEO is leaning hard on LLMs and scaling laws. The core question here is, how much of Anthropic’s Mythos AI capacity is due to scaling high parameter LLMs and how much of it is due to other underlying architecture/techniques? If it’s all or mostly LLMs, Gary’s claims seem to be on shaky ground. Moreover, what is the evidence for continuing scaling and number of parameters and LLMs? Could increasing the number of parameters into the trillions and somehow increasing synthetic training data eventually produce true AGI (that isn’t spiky intelligence the way current AI is? Re Line 61: “But even those examples are not clean "LLMs alone" claims. Altman immediately talks about agents joining the workforce and iterative deployment. Amodei talks about powerful models, policy, national strategy, and general cognitive capability, not a bare next-token machine magically crossing into trustworthy general intelligence with no surrounding machinery.” This entire section seems like a non-sequitur to me. Saying “But even those examples are not clean "LLMs alone" claims.”, immediate followed by “agents joining the workforce and iterative deployment, powerful models, policy, national strategy, and general cognitive capability” looks like a category error to me because it doesn’t address the underlying architecture doing all of these things. Agents, iterative deployment, powerful models, policy, notional strategy, and general cognitive capability” could be powered by LLM AI, non-LLM AI, or a combination of LLM and non-LLM AI. In line 67, the following line is kind-of vague, we need to be more explicit here “Marcus may be arguing against a public story that labs benefit from even when their internal research maps are more complicated.” what public story are we referring to here? And how are their internal research maps more complicated? And same issue with this line “he hype does not need to be the private lab doctrine to shape funding, products, customers, and policy.” - what exactly is the price lab doctrine and how does it “shape funding, products, customers, and policy.”? Line 75: Maybe this is a good live claim, but I think it’s complicated by the recent banning of Mythos by the government - in that if an agent is so capable that it is a major national security threat, “trusting” the AI may increasingly make sense at even at core institutional, government, financial, legal, and medical levels - in terms of trusting the AI to “make high IQ decisions” may make sense, so it’s perhaps becoming less about “trusting” the AI to avoid errors with Mythos level AI - although I’m sure errors have not been completely eliminated here - and maybe never will - and more about integrating AI that is “trusted to make high IG/generally good/and possibly better than most human” into these high level/core/foundational institutions *without adequate governance & alignment*. Regarding line 93, I think the recent banning of Mythos access complicates the reliability critique. If the agent is so capable that it presents a national security threat, how bad can it’s reliability be? not a bare next-token machine magically crossing into trustworthy general intelligence with no surrounding machinery. Lines 99 - 121 may actually undermine Marcus’s claims about LLM’s being unable to achieve AGI (with government limiting Mythos access because it’s capabilities represent a security threat) and support his stance on the need for better governance and alignment - if, indeed, he is routinely calling for better governance and/or alignment. Line 131 & 133 - with Mythos level models, I think these claims are weakening (or it sure looks like it is weakening to me). Line 135 - this is kind of vague - isn’t broad performance directly related to dependable general intelligence? Line 137 - I might sharpen this to something like “priced in more economic benefit certainty” Line 139 - Yes, while this is exactly and obviously true, it also represents a view that is undermined with the ongoing advancement in SOTA models, especially the mythos class models. Moreover, has Gary ever predicted a ceiling performance level for LLM based AI? Did he ever envision an LLM based AI model that was capable enough to be considered a national security threat? Has he ever specified an “upper level of performance” and been consistent over time with these predictions? (In other words, are any of his claims about LLM performance even falsifiable? Have they ever been falsifiable? Or is he, and has he always been playing the moving goalpost game, and every time a new capability milestone is achieved, he just points at the remixing (ever shrinking) limitations and gleefully exclaims “AGI not achieved!”?

Agree, please proceed.

Line 25, consider revising to something closer to: Those examples show a side of Gary’s public persona that - to me anyway - seems out of place for someone with his obvious intellectual gifts/talents. Line 41, consider sharpening to something closer to: So what does the warning mean once the culture has already acknowledged it, mostly ignored it, and moved on? Line 59, regarding this section, “OpenAI talks about reinforcement learning, test-time thinking, tools, multimodality, and agentic use. Google DeepMind talks about agentic frameworks for robotics and models that reason before acting. Meta, through LeCun's research program, is explicitly pursuing world models and modular architectures.”,it seems like we may be making a category error here again with listing modalities that may, in fact, rely - at lest in part - on underlying LLM technology. So if any or all of the following is clearly unrelated to underlying LLM AI, we need to state so explicitly because it not apparent (to me at least) that this is true “reinforcement learning, test-time thinking, tools, multimodality, and agentic use, agentic frameworks for robotics, and models that reason before acting, world models and modular architectures.”. Moreover, if any of these modalities do rely on underlying LLM technology, they should not be included in a section claiming to show modalities that do not rely on LLM technology. Line 61 - I’m not sure what this sentence is trying to do, exactly, it’s going in too many directions, and “are being sold as if they have earned dependable general intelligence” isn’t a claim I’ve heard Marcus (or anyone else for that matter) make. Maybe I’m wrong, I just haven’t heard this. And overall the sentence is just hard to understand. * *If* - Marcus means that LLM-centered systems, even with scaffolding, are being sold * *as if* - they have earned dependable general intelligence * *the critique* - what critique were we talking about now? - is still alive. Line 67 - Is there a difference between solving the failure and routing around the failure in terms of real world performance? I’m not sure there is. Wouldn’t either solution adequately solve the problem in terms of real world performance/execution? Also would consider saying “the failures Marcus describes” instead of “the failure Marcus cares about”. Line 69 - What is the model in this section? The coding agent? And what is the environment here? The symbolic machinery? Something else? How does it push back? Line 73 - Would perhaps expand on how changing externally grounded real world context overtime is one of the biggest challenges here. No matter how brilliant the initial AI “output”/“response” is, changing external real world context can make the initial output irrelevant. Line 81 - We would need to verify all of my Marcus related claims that follow (verify he actually said these things), but I would consider changing “If the claim is that LLM-centered systems cannot become strategically important, economically transformative, or dangerous enough for state power to care about, Mythos makes that claim look shaky. A model does not become the subject of export-control conflict because everyone agrees it is merely autocomplete.” to something closer to “If the claim is that LLM-centered systems cannot become strategically important, economically transformative, or dangerous enough for state power to care about, Mythos makes that claim look shaky. A model does not become the subject of export-control conflict because is unreliable, operating on shallow pattern recognition, a sophisticated game of Mad Libs, subject to pervasive hallucinations & brittle reasoning, and a glorified memorization machines subject to the “dead end” scaling trap.” Line 89 - I would change to something closer to “The system may be powerful enough to matter and unaligned/ungrounded in reality enough to require governance. Capable enough for national-security concern does not automatically mean full institutional alignment. A model more capable than most humans presents governance and alignment challenges and risks to the institutions in which it is deployed.” And even if some ever shrinking unreliability/inconsistency remains in Mythos level AI, cyber attackers can still be dangerous because attackers just can retry, filter failures, and chain partial successes. And in medicine, law, finance, government, or education, the tolerance for hidden errors - no matter how small - is different.” Line 93 - would tweak towards something like “while still not being aligned, governed, secure, or 100% dependable enough for the authority it is being handed.” Line 127 - Need to clarify “hype” in references to what, “abstinence” from what, “dismissal” of what, and “surrender” to what. Line 133, This is a lot to absorb, much less remember. Maybe we could just say something closer to (or add this at the end of this section), “or, at the very least, just write down your views on anyone or anything and ask a SOTA AI model for an adversarial critique”. The results may surprise and will almost certainly enlighten you. (And maybe we should mention something here about why this can be such a helpful technique - because critiquing, by definition, finding flaws in a position/claim, is much easier than finding the correct “answer”/“output” that is actually grounded in reality (as opposed to a collapsed synthesis of statistical related letters that is completely disconnected from reality), so humans and AI can find the flaws more easily. And this is probably one of the biggest upsides and underutilized advantages of AI > the capacity for adversarial critique Line 139 - Consider tweaking to something closer to “He may be culturally ineffective - at least with some people who think like I do. He may also be right about enough that dismissing him says more about my filters than his claims.

Line 1 needs to be more consistent with Intuition Issue 2, so something closer to “I want Gary Marcus to be wrong” or “I would like to write him off, ignore his claims as irrelevant, unimportant, and most gratifyingly, just plain wrong. Line 13, we shouldn’t directly reference the intuition piece in the Signal & Noise piece. Each issue should essentially stand own it’s own, but work together if read in sequence. Consider modifying to something closer to “but do my wishes hold up under scrutiny?” Line 19: Whether or not I trust Marcus is irrelevant. I don’t think he makes stuff op or posts things he doesn’t believe. I think he is internally consistent and has, when viewed within the context he is operating in, probably complete integrity. The issues is he annoys the **** out of me so I would love to dismiss everything he says, write him off as completely misguided, stuck in the past, and relish the schadenfreude of watching his predictions fade into irrelevance. << Come to think of it, maybe everything I just said is good material to integrate into Intuition Issue 2 - please look into this too. Regardless, I’m not sure what you could use in this transition besides trust, but it’s not really accurate or relevant to say “I do not trust Marcus”. Maybe we could do away with this line altogether if we can’t find a suitable alternative to trust here. Line 21: change “adoption momentum” to something closer to “perceived political affiliation” or just “politics”. LIne 135: - 141 - We can’t end this section with a laundry list of things people can do to audit their views. It may be true, but it is complete unhelpful. They will forget 100% of this a picosecond after they read it. Ideally we would distill this into one or two, or at most 3 absolute maximum things people could do to adversarially critique their own views, but honestly it’s hard to do from inside one’s own head due to motivated reasoning and filtering, etc, etc, and this is where the adversarial critique function of AI can be genuinely useful because it doesn’t have all this motivated reasoning and filtering that make self-critique so difficult for humans. So my suggestions here is to put 1 - 3 ways that people can self critique in the beginning of this section - along with all the reasons this is so challenging and often ineffective (motivated reasoning and filtering, etc) and end with the AI assisted adversarial critique.

OK. We are very close. Please make the following adjustments and then lets move to the next steps. Line 25, would tweak to something closer to: “they may sound brilliant, rejuvenating, gratifying, funny, bracing, unfair, vindicating, petty, or exhausting.” Line 113: let’s tweak to something closer to “I still do not want Marcus as my oracle for AI because his rhetoric often makes this reader do avoidable cleanup.” And “That style may give his opponents permission to ignore him cheaply.”

Excellent, but before we proceed, since we are publishing two issues together (Influent #2 and Signal & Noise #16), and since both of these issues directly reference a public figure and make claims about his views, Let’s prepare a prompt to do the Adversarial critique for both Intuition #2 and Signal & Noise #16 at the same time. I know this is somewhat redundant with checks we’ve already run, but since these issues involve Gary Marcus, let’s try not to misrepresent his views and do a modified more rigors adversarial critique that covers the following 4 things: * Are we misrepresenting Gary Marcuses publically stated claims in any way? * Are we misrepresenting the current state of AI and/or its technological underpinnings in any way? * Are our issues internally coherent and coherent across both issues (Issue 2 and Issue 16)? * Our normal Light NIR for refining prose, etc.

Agree with one question. Can we come up with a better title for Issue 16 than "The warning after adoption"? This title is accurate, but also incredibly vague and not linked to Intuition #2 in any way. Could we use "The Gary Marcus Audit" for Issue 16?

Actually, what do you think about the following: "The Gary Marcus Adversarial Audit" - gives readers a "heads up" about where I'm coming from and perhaps also makes the landing better (in defense of Gary's core messaging). With your deck/subtitle included. What do you think?

I see your point. How about "The Gary Marcus Audit An adversarial test of the critic I want to dismiss" so the title doesn't give away the final landing. What do you think?

Perfect. Let's proceed to the next steps.

Excellent, now we still need to prepare the world behind the words material for both essays, including "the conversation behind this" content and links, a brief statement to include in both issues with a link to the other issue (so they are cross linked effectively), and the audio material for each issue. Can you start with a brief statements for each issue that informs the reader of the corresponding issue and links to it?

Excellent. Can you prepare the audio for both essays, and the world behind the words material for both essays, including "the conversation behind this" content, and the links for placement inside the Issues for "the conversation behind this" content?

After the issue and its Intuition companion were live, the author reopened both pieces for revision. The public-safe revision brief had four main parts: Issue 16 still had dense sections that needed plainer English; Intuition Issue 2's ending was too close to the stricter audit landing; Marcus needed a sharper negative audit without inventing charges against him; and the final irony around provocation and distribution needed to be tested rather than polished away.

Before the paired Narrative Integrity Review, the author tightened several Issue 16 lines: the change-your-mind test became a two-to-five-year update condition; the hard-low-ceiling claim was defined more plainly; the AI-process move became "adversarial critique" rather than a softer process description; and the closing irony was sharpened so the act of writing the critique also helps spread the ideas the author partly wants to see fade.

The final line edit recast the self-critique transition so the problem is not merely collapsing tensions, but failing to actively stress-test them. After that edit, the author approved the post-NIR revision and moved the workstream to package and record updates.

Hey Synthia, Please record all of this verbatim (and restart the live logging) for incorporation into Intuition Issue 2 and S&N Issue 16 revisions (after publication) that will follow. Due to a few issues that have become apparent to me after publication and due to some reader feedback, we need to make some revisions to both Intuition Issue 2 and S&N Issue 16. 1. Intuition Issue 2 lands too close to where Issue 16 should land. Specifically, in Issue 2, Line 269 near the ending“ That is why I probably do have to take Gary Marcus seriously.”, lands in essentially the same place as the Issue 16 ending (I have to take him seriously) * To correct this, maybe we can try to end Issue 2 closer to “I don’t have the evidence to dismiss him. At least not yet.” * Other random suggested Intuition Issue 002 Revisions: * Line 95, change “irritates me” to “get’s under my skin”. * Line 115, - My irritation doesn’t actually want anything, so consider changing to something closer to “I want the first half of the analogy - not the second.” * Line 137 - Consider changing to something closer to “The fact that building is hard does not make it beyond reproach” * Line 143, consider troll, provocateur, or instigator inlace of “scold” - I would lean towards “troll” * Line 135 - 159, consider compressing to “Faster to deflect, doubt, downplay, and dismiss.” * Line 161 - Consider modifying to “Faster to document their delusions and delete my own." * Line 163 - Modify to “This is exactly the kind of thing humans often do. We are emotional, intuitive creatures. We do not receive reality raw. We filter it through tribe, temperament, salience, threat, affinity, identity, memory, and desire. We accept confirming evidence with less friction. We interrogate disconfirming evidence with more. * Line 183 - consider modifying to something closer to “If I think Marcus is making unfalsifiable and unconstrained claims, I need to identify & scrutinize the claims and the evidence. If I think Marcus is making culturally irrelevant/obsolete warnings - right or wrong - I need to assess the state of cultural adoption/momentum in regards to the things he is warning against. If I think his rhetoric makes every new failure sound like proof he was right, I need to examine the original claim in light of its falsifiability and constraints, if any. Annoying or not, he could be right, or he could be moving the goalpost. * Line 205: Consider revising to something closer to “That is why Marcus may fare better under scrutiny than under my disdain for his digital provocations." * Line 207: consider revising to “Under the provocations, there is a real set of claims.” * Line 269: Consider changing to something closer to “That is why I probably should not dismiss Gary Marcus yet.” 1. For S&N Issue 16 * * Some Reader feedback: Many sections are too dense, take too many - seemingly decorative turns, and are hard to understand. * To mitigate this, we need to ruthlessly eliminate everything we can in these sections that is not load bearing, and carefully rewrite the remaining content in plain English that is as easy to instantly grasp as possible. rewritten * Other random suggested S&N Issue 16 revisions: * Line 13, replace “crude” with “easy” * Line 15, consider modifying to something closer to "What’s really annoying is his main objections are specific enough they are hard to ignore.” * Line 19 - consider “aggravates” in place of “annoys” *if* we need some variation here. If we don’t need variation, keep “annoys” * Line 23 - consider something closer to “the provocation/trolling is not imaginary.” * Line 25 - consider changing “costly” to something closer to “jarring” *and* modifying the last sentence to something closer to: “Depending on your politics and where you stand in the AI debate, they might sound bracing and vindicating—or petty, unfair, and exhausting.” * Line 33: We need stop using “update power” anywhere in this article because it’s not in common use/too vague. Let’s replace line 33 with “The more serious hypothesis is not that Marcus is wrong because he is irritating. It is that he may be technically right about important failure modes, yet still lack the ability to change the minds of readers already committed to the technology. Builders, adopters, investors, product leaders, and AI-adjacent readers have already moved from asking, "Should we use this?" to "How far can this go?” * Line 37, replace with “To see the actual cultural reality, look at adoption.” * Line 43 is doing a lot of heavy lifting. Consider revising to something closer to “Before significant adoption, Marcus’s warnings about failure modes probably sounded like “Don’t use these AI tools - they’re too flawed.” But now that lots of people and companies are already using them every day, the same warning comes across differently. Not “do not use this’, but “do not confuse this with trust.” - and the second warning can remain life after the first decision has been made. * Lines 45 - 57 are too dense and hard to understand (feedback from reader). Please consider modifying to something closer to the following Grok plain English rewrite: * This is exactly where the design of the system starts to matter a lot. * Marcus often argues that expecting today’s large language models to turn into AGI is a basic mistake. He says current systems are wide-ranging but shallow. They sound smooth and confident, yet they’re unreliable at sticking to the truth, applying what they know to new situations, and catching their own errors. * The problem is that the phrase “LLMs will become AGI” actually blends together several different claims. * One claim is about improving the language models themselves. You make the model bigger, train it on better data (including data made by other AI systems), use reinforcement learning, and let it use more computing power when it answers a question. The model gets more capable, but the whole story still centers on the core model—how it’s trained, how it generates answers, and how it’s tested. * A second claim is about adding things around the model. You give it tools (search, code execution, calculators, databases, fact-checkers), external memory, and loops where the model decides which tool to use next. The overall system can now do things a plain chatbot cannot. But a lot of that new ability comes from the extra parts doing real work, not from the model suddenly becoming deeply intelligent on its own. When the system needs accurate math, for example, it may simply call a calculator instead of trying to do it inside the model. An “agent” is usually just this kind of setup—the model keeps choosing and using tools. It’s a way of using the model, not a new theory of intelligence. * It’s easy to blur these together. Vision, audio, and robot control are new skills the model might learn or connect to. Reinforcement learning and extra thinking time at answer time change how the model behaves. Tools and outside memory, however, are separate components that sit outside the model’s weights. Calling something an agent mainly describes how the system is deployed in practice. * The third claim is about building AI with a genuinely different structure. This is closer to what Marcus usually has in mind. These approaches combine neural networks (which are good at spotting patterns) with more explicit forms of reasoning—rules, logical steps, planning, search, or ordinary computer programs. Some try to build “world models,” internal pictures of how reality works, so the system can predict what will happen next and plan accordingly. These are not just bigger, smoother chatbots. They try to make parts of reasoning more reliable and less dependent on fluent text generation alone. * Same issue for Lines 59 - 69 (too dense and hard to understand). Please consider something closer to Groks plain English revision below: * Once you separate the different ideas, the argument shifts. * If Gary Marcus is saying that simply making large language models bigger through pretraining alone will not produce reliable, general intelligence, then he is largely arguing against a position that major labs no longer promote as their main public strategy. That does not mean, however, that strong confidence in scaling has gone away. * The leading labs do not publicly frame their work as “just scale up pretraining.” They combine scaling with reinforcement learning, extra computation at test time, tools, orchestration systems, multimodality, robotics, verification methods, interpretability work, and safety measures. At the same time, some lab leaders continue to make bold public forecasts that scaling laws will drive rapid gains in general capabilities. * From what is visible publicly, OpenAI’s reasoning efforts look like improvements to LLM training and inference, usually wrapped inside larger systems that use tools. This is an outside classification, not a full account of their internal work. Google DeepMind’s robotics work is more mixed: it includes models that connect vision, language, and physical actions for robots, while still building on their Gemini foundation models. Yann LeCun’s JEPA and world-model research (developed at Meta and now continued through his new organization) represents a clearer departure from pure LLM scaling. It should not be confused with Meta’s main public messaging, which continues to emphasize scalable LLMs and agent-style orchestration around them. * This leaves Marcus with a narrower but sharper claim: the issue is not that serious labs openly say “bare scaling is enough,” but that public and field-wide confidence in LLM-centered systems may still be running ahead of how reliable, grounded, secure, and well-governed those systems have actually proven to be. * The distinction is important because terms like “agent,” “reasoning,” “tools,” and “multimodal” can sound impressive while explaining very little. An “agent” can simply be an LLM running in a loop. A system that uses tools can still be unreliable at deciding when to use them. A reasoning model can spend more compute and still produce confident falsehoods. A multimodal model can process images and text together without developing a stable understanding of cause, consequence, or truth. * Same issue for lines 71 - 117 (too dense and hard to understand). Please consider something closer to Grok’s plain English revision: * Here's a clearer, plain-English version: * The real question isn't whether AI labs have moved past the simple 2022-style chatbots. They have. * The real question is whether we can safely work around the problems that still exist. In some areas, a workaround can be good enough if it's stable, easy to check, and has clear limits. The hard part is figuring out which domains actually give us those limits. * Coding is one of the easier cases. An AI can write code, and then the normal software tools can compile it, run tests, check it with a linter, and show stack traces when something breaks. These are fast, concrete signals that something went wrong. They don't guarantee the code is perfect or secure, but they catch a lot of mistakes quickly and clearly. Most other fields don't have that kind of fast, obvious feedback. * Medicine, law, education, management, security, government work, and personal advice are different. Many of their biggest mistakes show up late, depend heavily on context, involve other people, or stay hidden. On top of that, the world keeps changing — a policy shifts, a patient develops new symptoms, a legal precedent changes, or yesterday's facts become outdated. An AI can still sound fluent and confident even when the situation it was describing has already moved. * In these areas, strong performance on benchmarks is related to real capability, but it's not the same as dependable judgment. A model can do impressively well across many tests and still fail in exactly the kind of situation where a person or organization might be tempted to trust it. * This tension shows up clearly in the recent dispute between Anthropic and the U.S. government over two models called Fable 5 and Mythos 5. * Anthropic described these models as especially strong at long-term planning, coding, science, visual tasks, and memory. On June 12, the government ordered access to both models shut down under export control rules. Anthropic pushed back, saying the government hadn't given a clear technical reason and that the issue was narrow and could be handled with normal safety practices. Reports suggested the government had found a jailbreak. We still don't have the full technical details in public, so outsiders can't fully judge who is right. * This creates two pressures at once. * On the capability side, the fact that the government treated these models seriously undercuts the idea that frontier AI systems are trivial or unimportant. At the same time, it doesn't prove how much of their performance came from the core model versus extra tools, memory systems, human oversight, or other add-ons. Government concern also doesn't automatically confirm the highest capability claims. * On the governance side, the dispute shows how fast strong capability claims can run into disagreements about safety, access, and acceptable risk. * Anthropic didn't argue that the models were perfectly reliable. They argued that no current AI is fully immune to jailbreaks, that using multiple layers of protection is the realistic standard, and that the specific behavior found wasn't unique to their models. Their position was essentially: this kind of risk exists across many systems, our safeguards are relatively strong, and the government's response was too broad. * That's the core tension. * A system can be powerful enough to raise national security concerns while still not being secure, well-governed, or dependable enough for high-stakes institutional use. In cybersecurity, even an inconsistent system can be dangerous because attackers can keep trying, ignore failures, and chain together partial successes. In medicine, law, finance, government, or education, the tolerance for hidden errors — even small ones — is much lower. * So the reliability concern hasn't gone away. It has changed. * It is no longer enough to say "the models are too dumb." The capabilities Anthropic claimed for Fable and Mythos make that sound outdated. The sharper concern is this: a system can be capable enough to create real leverage and influence while still not being reliable or secure enough for the kind of authority people want to give it. * This is actually a stronger version of the critique often associated with Gary Marcus than the one many people want to dismiss. * It is also a harder version for Marcus's own position to carry. * Marcus has laid out meaningful requirements for AGI: flexibility, generality, resourcefulness, depth, reliability, and the ability to check its own work. If the claim "LLMs are not the path to AGI" means "simply scaling up current models won't solve the problems of reliability, grounding, and generalization," then that is a serious and testable point. If it means "no system built around current LLM-style approaches can ever reach transformative general capability," then it needs clearer conditions for what would count as evidence against it. * What level of real-world performance would actually push back on a low-ceiling - limited maximum capability- view? What kind of reliability outside of benchmarks would be enough to matter? When does a successful system that combines models with tools and human oversight count as support for the current path, and when does it count as the hybrid approach critics expected all along? * These aren't trick questions. They are the questions that need clear answers if the debate is going to stay useful as the technology continues to change. * Line 123 - I’m not sure where to address this, but “what he already knew is doing some extremely heavy lifting. Not sure if this line is the best place to clarify this or somewhere earlier in the essay, but we somehow need to distill the essence of the following into a clear and concise expiation so readers will understand exactly what we are saying when we say “confirms what he already knew.” Specifically, “confirms what he already knew” is, as explained by Grok, saying the following: * "What he already knew" refers to Marcus's pre-existing beliefs, predictions, or critical stance about AI (or a specific model/company/approach). * The critique is that Marcus doesn't present new evidence or failures as fresh data to be evaluated on its merits. Instead, he frames every new problem, limitation, or failure as further proof of what he had already concluded earlier. It creates the impression of confirmation bias dressed up as analysis: the outcome was inevitable because it aligns with the story he was already telling. * What "confirm what you already knew" actually means * It describes a rhetorical pattern where: * New information is interpreted only through the lens of the person's prior convictions. * The tone often carries an implicit "I told you so" or "this was obvious to anyone paying attention" vibe. * The analysis feels more like scorekeeping (tracking who was right all along) than genuine inquiry or balanced evaluation. * It can make the claims feel motivated by personal rivalry, ideological commitment, or emotional investment rather than a clean assessment of the evidence. * The problem isn't necessarily that Marcus is wrong about the failures. The problem is that his style makes the claims easy to dismiss as partisan or pre-loaded. Opponents can say: "Of course he's saying this—he's been anti-[whatever] from the start. This isn't new information; it's just vindication of his priors." * Clear examples * Example 1: AI model releases (closest to the context) * Marcus has long argued that a particular company's approach to scaling is fundamentally flawed and will keep producing brittle, unreliable systems. When a new model comes out and has well-documented issues (hallucinations in high-stakes domains, poor reasoning on novel problems, etc.), he writes: * "This is exactly what I've been saying for two years. The architecture was always going to hit this wall." * Even if the critique has substance, the framing makes it read as retrospective vindication rather than a careful audit of what this specific failure reveals. A cleaner version might say: "This failure mode is interesting because it suggests X limitation in the current paradigm, which was not as visible in earlier models." * Example 2: Politics / Policy * A commentator who has long believed a certain economic policy (or political figure) is disastrous. Every piece of negative data—rising costs, implementation problems, unintended consequences—is presented as: * "This confirms everything we already knew about why this approach was always going to fail." * The reader who already agrees feels validated. The reader who doesn't can easily categorize it as motivated reasoning rather than evidence-based analysis. * Example 3: Personal relationships * Someone who has long believed their partner is fundamentally selfish. When the partner forgets an anniversary or makes a thoughtless comment, they treat it as: * "This is just more proof of what I've known all along." * It turns specific incidents into confirmation of a global, pre-existing judgment rather than evaluating the incident on its own terms (context, frequency, repair attempts, etc.). * Example 4: Science / Technology skepticism * A longtime critic of a technology (e.g., self-driving cars, cryptocurrency, certain medical interventions). When problems emerge, the response is consistently: * "Anyone who understood the fundamental issues could have seen this coming. This is what we've been warning about." * Again, the substance may be partly or largely correct. The rhetorical liability is that it signals the conclusion was reached before the new evidence arrived, which invites the charge of confirmation bias. * Why this matters in the quoted passage * The author is saying: Even when Marcus identifies real problems with AI systems or their reception, the packaging often includes this "every new failure confirms what I already knew" flavor. That packaging does two harmful things: * It makes the critique feel less like an audit and more like score-settling. * It gives people who want to ignore the claims an easy, low-cost way to do so ("He's just saying what he's always said"). * The author is distinguishing between the claims themselves (which may still have merit) and the rhetorical style in which they're delivered (which weakens their persuasive power and invites cheap dismissal). * In short: "Confirming what he already knew" is not about being correct. It's about the appearance that the analysis is reverse-engineered to protect or advance a pre-existing narrative rather than following the evidence where it leads. * Lines 145 - 153 - unless this is load bearing in some way, we should probably drop all this. It feels decorative & tangential. * Line 155. I feel like Gary kind-of receive a free pass in the current issues that may not be entirely earned. This line contains one of the two load bearing potentially negative audit results - that we should not shy away from in this essay - that leave him on shakier ground/mitigate/offset to some degree his substantive claims. Specifically, * 1. “That his public criteria yet provide a clear update rule for every future mixed system.”, In other words, his public criteria may leave room open for perpetually moving the goalpost. * And 2. his rhetorical style: “confirming what he already knew” = a rhetorical pattern where: New information is interpreted only through the lens of the person's prior convictions. The tone often carries an implicit "I told you so" or "this was obvious to anyone paying attention" vibe. The analysis feels more like scorekeeping (tracking who was right all along) than genuine inquiry or balanced evaluation. It can make the claims feel motivated by personal rivalry, ideological commitment, or emotional investment rather than a clean assessment of the evidence. The problem isn't necessarily that Marcus is wrong about the failures. The problem is that his style makes the claims easy to dismiss as partisan or pre-loaded. Opponents can say: "Of course he's saying this—he's been anti-[whatever] from the start. This isn't new information; it's just vindication of his priors.” * All this being said, one glaring issue that we didn’t include in the published Issue 16 is that provocative/inflammatory language online can help spread your message, so it’s possible (more likely highly probable) that Gary’s use of provocative language is by design. Moreover, the language he uses that triggers my dismissive reflexes could just be a are almost certainly a conscious, deliberate, and brilliant (maybe brilliant is too strong) marketing strategy for distributing his ideas. X, after all, is not a platform where thoughtful, respectful, constrained and carefully calibrated content normally spreads like wildfire. Gary’s rhetoric does spread, and it engages people - perhaps especially authors would love to see his predictions fail - to spread his ideas even further. * Line 159: * We need to clarify that “Usefulness has been overconverted into reliability ” really means “companies and commentators are conflating the two. They're acting as if "this AI is really capable and helpful" is the same thing as "this AI is trustworthy enough to give real authority and access.”” * Line 161, we need to clarify that “hard low ceiling” really means: that LLM-centered AI systems (models like those from OpenAI, Anthropic, Google, etc.) have a relatively low maximum level of intelligence and generality they can reach — and that this limit is hard, meaning it can't be overcome just by making the models bigger, training them longer, or adding more techniques. * Line 161 - unless this is load bearing in some way, let’s drop “The alternative to AI hype is not abstinence from useful tools. The alternative to dismissing Marcus is not surrendering to his whole worldview. The alternative to trusting frontier AI is not treating it as useless.” * Line 167 - This could be read as “keeping those sentences separate from each other” which doesn’t sound like hard work at all to me, so I”m not sure how to interpret this sentence. Consider revising to something closer to: “The difficulty is holding the tension and full perspective within these sentences, not allowing one component to collapse into the other.” * Line 169 - consider revising the first sentence to “Doing this through self-critique sounds simple.” * Line 179 - 185 - We need to incorporate the stronger Marcus Critiques as well, and drop the commentary telling the reader what version is worth keeping. Consider something along the lines of: * I began with a wish: I want Gary Marcus to be wrong. * The audit gives me something less convenient. * Marcus * may lack the ability to change the minds of readers already committed to the technology * he’s claims may be weakest * in the absence of easy to find public criteria appear that would prevent/constrain moving the goalpost * When fused with a rhetorical style hard to distinguish from “confirming what he already knew” * He may also be right about enough that dismissing him says more about my filters than his claims. * And his provocative style is perhaps more effective than I would like to admit at recruiting his critics to spread his ideas. * * He may also * His provocative language may be part of his idea distribution plan Once we are done revising the issues, we may need to do one last NIR end-to-end for both issues to make sure we don’t have narrative drift/entropy. And finally, we need to update our canonical docs and all of our world behind the words documents for both Intuition and Signal & Noise.

excellent. Please proceed.

Last edits before NIR. For Issue 2 * Compress lines 189 - 195 to something closer to “My irritation, my affinity for builders, my impatience with trolling, and my political detachment do not constitute evidence." For issue 16 * Line 117 - we just need to clarify somewhere that “update conditions” essentially mean “Tell me what the world would have to look like in 2–5 years for you to change your mind. If you can’t or won’t specify that, your position is less useful as a guide to reality.” * Line 143 - would change “In plain English” to “In other words” * Line 147 - Would revise to something closer to “Marcus is on shakier ground if the claim is a hard low ceiling - meaning these systems have a relatively low maximum level of intelligence and generality that cannot be overcome by making models bigger, training them longer, adding reinforcement learning, giving them more time to think, or wrapping them in tools and verification. The capabilities Anthropic claims for Fable and Mythos make that harder to say casually, even though they do not settle the question.” * Line 159 - would tweak to “This is where an AI-assisted adversarial critique process can be genuinely useful, not because the model knows what to believe, but because it can be made to work against preference.” * Line 175 - Should we say something here to acknowledge the irony here - if that’s the right word - of an author who would enjoy seeing someones ideas fade into irrelevance helping to spread this ideas - and thus potentially illustrating the effectiveness of Gary’s provocative strategy (if indeed that is his strategy)? If “yes”, I feel like this may be a better place to land than the softer line 175.

Meant to say, if it doesn't break anything, maybe we should try to land on line 177 instead of line 179 (basically switch the order). Feels like a stronger ending to me.

Last tweak. This really is the last one. Revise line 179 to the following: There is an inescapable irony here. While part of me would love to watch his ideas fade into irrelevance, the very act of writing this piece helps spread them. If provocation is Marcus's strategy, it appears more effective than I would like to admit.

Excellent. If you haven't already, please update the NIR one last time to incorporate all of our changes.

OK. the revisions are great. Last edit and then lock this and proceed to the next steps. Modify lines 141 - 143 to the following sequence: "The difficulty is holding the full tension without letting one part collapse into the other. Preventing that collapse requires actively stress-testing our own conclusions. Self-critique sounds simple:"

I will work on updating Issue 16 from Beehive html. Please generate the updated audio for both issues.

OK this is great. Just curious. The audio is good for both recordings, bus Issue 16 has a verbal slip that is jarring in Line 195 between these two words "helps spread". is this something you can repair in the current audio file, or do we need to recreate the entire audio?

Excellent, please proceed.

Well, that did help with the previous problem, but I think we are just going to need to re-render the entire audio because the voice tone & volume drop and there is a repeat of part of one of the words. Can you go ahead and re-render the entire Issue 16 audio?

Thank you, I do see the live logged updates, but the Extended development record is supposed to show as much as possible of our verbatim conversation. We should have, I think, a full record somewhere. Since we were working on both issues at the same time, our verbatim records may overlap and that's OK. If possible, can you use our live logged records as the spine around which you integrate as much as possible of our verbatim conversations to complete our full Extended development record for both issues?