I like this but at the same time it seems tricky don't you think? Is the AI model intuiting your intent? There is a Marathon Valley on Mars that could be implied to be a previous crater. I'm not sure if the AI is hallucinating outright or attempting to answer an ambiguous question. It's like saying "tell me about the trade building in New York". Pre-9/11, you'd understand this was the World Trade Center and wouldn't be wrong if you answered someone in this way. "Tell me about the Triangle statue". "Oh the Triangle statue was built in ancient egypt around BC 3100". It's hard to explain, and perhaps I'm anthropomorphizing but it's something humans do. Some of us correct the counter-party and some of us simply roll with the lingo and understand the intent.
Yes, I understand the basics. LLMs predict the next most probable tokens based on patterns in their training data and the prompt context. For the 'Marathon crater' example, the model doesn't have a concept of 'knowing' versus 'not knowing' in our sense. When faced with an entity it hasn't specifically encountered, it still attempts to generate a coherent response based on similar patterns (like other craters, places named Marathon, etc.).
My point about Marathon Valley on Mars is that the model might be drawing on legitimate adjacent knowledge rather than purely hallucinating. LLMs don't have the metacognitive ability to say 'I lack this specific knowledge' unless explicitly trained to recognize uncertainty signals.
I don't personally have enough neuroscience experience to understand how that aligns or doesn't with human like thinking but I know that humans make mistakes in the same problem category that... to an external observer.. are indistinguishable from "making shit up". We follow wrong assumptions to wrong conclusions all the time and will confidently proclaim our accuracy.
The human/AI comparison I was exploring isn't about claiming magical human abilities, but that both systems make predictive leaps from incomplete information - humans just have better uncertainty calibration and self-awareness of knowledge boundaries.
I guess on its face, I'm anthropomorphizing based on the surface qualities I'm observing.
Okay but by your own understanding it's not drawing on knowledge. It's drawing on probable similarity in association space. If you understand that then nothing here should be confusing, it's all just most probable values.
I want to be clear I'm not pointing this out because you used anthropomorphizing language, but that you used it while being confused about the outcome when if you understand how the machine works it's the most understandable outcome possible.
That's a fair point. What I find interesting (and perhaps didn't articulate properly) isn't confusion about the LLM's behavior, but the question of whether human cognition might operate on similar principles at a fundamental level - just via different mechanisms and with better calibration (similar algorithm, different substrate), which is why I used human examples at the start.
When I see an LLM confidently generate an answer about a non-existent thing by associating related concepts, I wonder how different is this from humans confidently filling knowledge gaps with our own probability-based assumptions? We do this constantly - connecting dots based on pattern recognition and making statistical leaps between concepts.
If we understand how human minds worked in their entirety, then I'd be more likely to say "ha, stupid LLM, it hallucinates instead of saying I don't know". But, I don't know, I see a strong similarity to many humans. What are weight and biases but our own heavy-weight neural "nodes" built up over a lifetime to say "this is likely to be true because of past experiences"? I say this with only hobbyist understanding of neural science topics mind you.
How do they work? My understanding is each 5 characters are tokenized and assigned a number. If you take gpt2 it has 768 embedded dimensional values which get broken into 64 which creates 12 planes. When training starts random values are assigned to the dimensional values (never 0). Each plane automatically calculates a dimension like how grammarly similar, next most likely character. But it does this automatically based on feedback from other planes. That's where I get lost. Can you help fill in the pieces?
Large corporations aren't demonstrably harmed by DEI initiatives, yet all hiring processes should be merit-based. Encouraging more women in engineering is a reasonable goal, but lowering qualification standards would be problematic. Throughout my career, I've never observed standards being lowered for diversity candidates—though I've witnessed DEI being wrongly blamed when underperformers are terminated.
What I have consistently seen is hiring standards fluctuating based on company performance: loosening when profits are high, tightening when they're not. The most pervasive bias in hiring isn't DEI-related but rather social network preference, where managers favor friends, neighbors, or people similar to themselves regardless of qualifications. This mirrors the "backdoor" admissions seen at elite universities and extends to government appointments, where connections often appear to outweigh merit...
Here's a plausible steelman argument, drawing from insights from CEA Chair Stephen Miran and a paper he authored pre-administration titled "A user's guide to restructuring a global trading system"
It is a blueprint for "tariff based upheavel". It proposes using 'unilateral U.S. tariffs as leverage' to force other countries into a new accord, dubbed by some a potential "Mar-a-Lago Accord," analogous to Bretton Woods, that would include 'coordinated currency realignments'.
Miran argued that because the strong dollar has made U.S. exports uncompetitive and fueled chronic deficits, the U.S. might need to pressure other countries to strengthen their currencies (i.e. weaken the dollar) through a trade war if necessary. Alongside this potential grand strategy, it's also argue the tariffs directly address specific issues like unfair trade practices (the 'reciprocity' argument), dumping, reliance on adversarial supply chains (national security), and aim to incentivize domestic manufacturing investment.
Elements of Miran's thinking in the paper are evident in the Administration's approach. For instance, Trump aides publicly claim they are targeting countries with "artificially devalued" currencies for tougher tariffs and the President frames tariffs as "reciprocal", a hint that the endgame is to make others lower their trade barriers or adjust currency values to balance trade.
Direct quote:
"Treasury Secretary Scott Bessent has mentioned, the US could have a “traffic light” system that divides the world into three tiers: “green” countries with shared values, aligned economic and security goals, and a willingness to cooperate on exchange rates; “yellow” or neutral countries that want to keep high tariffs and remain outside the US defense system; and “red” countries, meaning adversaries or sanctioned nations that refuse to cooperate."
I think that adds potential useful context around where we might expect countries to align and what the intent is. Or perhaps as I've seen it put more inflammatorily, there are "vassals", "neutral" and "adversaries".
One last note would be that comparisons to the 1930s Smoot-Hawley tariffs are often made but the context is fundamentally different. In the 1930s, the US was a major creditor nation with large trade surpluses and dominant manufacturing. Today it's a large debtor nation seeking to revitalize its industrial base and reduce deficits within a far more globalized system. It's not quite fair to call it protectionism (as a sole objective) and the time period doesn't generalize to today's America.
With those datapoints....
The steelman perspective is that this is less short-term theater and more a high-stakes, potentially disruptive strategy aimed at fundamentally restructuring the global trading system. The intended endgame is to reassert U.S. economic advantage, enhance national security through resilient supply chains, and better align global economic rules with U.S. interests in a changed geopolitical landscape, using U.S. market access and currency centrality as key leverage.
To be clear, this doesn't imply that I agree. I’m not convinced it’ll succeed as intended, but that’s the best-case rationale. It might be a "Hail Mary" to prevent ceding global leadership to China.. or a way to "hit reset".
I love this. The more people that say "I don't get it" or "it's a stochastic parrot", the more time I get to build products rapidly without the competition that there would be if everyone was effectively using AI. Effectively is the key.
It's cliche at this point to say "you're using it wrong" but damn... it really is a thing. It's kind of like how some people can find something online in one Google query and others somehow manage to phrase things just wrong enough that they struggle. It really is two worlds. I can have AI pump out 100k tokens with a nearly 0% error rate, meanwhile my friends with equally high engineering skill struggle to get AI to edit 2 classes in their codebase.
There are a lot of critical skills and a lot of fluff out there. I think the fluff confuses things further. The variety of models and model versions confuses things EVEN MORE! When someone says "I tried LLMs and they failed at task xyz" ... what version was it? How long was the session? How did they prompt it? Did they provide sufficient context around what they wanted performed or answered? Did they have the LLM use tools if that is appropriate (web/deepresearch)?
It's never a like-for-like comparison. Today's cutting-edge models are nothing like even 6-months ago.
Honestly, with models like Claude 3.7 Sonnet (thinking mode) and OpenAI o3-mini-high, I'm not sure how people fail so hard at prompting and getting quality answers. The models practically predict your thoughts.
Maybe that's the problem, poor specifications in (prompt), expecting magic that conforms to their every specification (out).
I genuinely don't understand why some people are still pessimistic about LLMs.
Great points. I think much of the pessimism is based on fear of inadequacy. Also the fact that these things bring up truly base-level epistemological quandaries that question human perception and reality fundamentally. Average joe doesnt want to think about how we dont know if consciousness is a real thing, let alone determine if the robot is.
We are going through a societal change. There will always be the people who reject AI no matter the capabilities. I'm at the point where if ANYTHING tells me that it's conscious... I just have to believe them and act accordingly to my own morals
This article does not line up with my experiences at all. Sometimes I wonder if it's something to do with prompting or model selection.
I recently built out a project where I was able to design 30+ modules and only had 4 generation errors. These were decent size modules of 700-5000 lines each. I would classify the generation errors as related to missing specification -- i.e., no you may not take an approach where you import another language runtime into memory to hack a solution.
Sure, in the past, AI would lead me on goose chases, produce bad code, or otherwise fail. AI in 2025 though? No. AI has solved many quirky or complex headscratchers, async and distributed runtime bugs, etc.
My error rate with Claude-3.7-sonnet and OpenAI's O3-mini has dropped to nearly zero.
I think part of this is how you transfer your expert knowledge into the AI's "mindspace".
I tend to prompt a paragraph which represents my requirements and constraints. Use this programming language. Cache in this way. Encrypt in this way. Prefer standard library. Use this or that algorithm. Search for the latest way to use this API and use it. Have this API surface. Etc. I'm not particularly verbose either.
The thinking models tend to unravel that into a checklist, which they then run through and write a module for. "Ok, the user wants me to create a module that has these 10 features with these constraints and using these libraries."
Maybe that's a matter of 25yrs of coding and being able to understand and describe the problem and all of its limits and constraints quickly but I find that I get one-shot success nearly every time.
I'm not only laying out the specification, but I also have the overall spec in my mind and limit the AI to building modules to my specifications (apis/etc) rather than trying to shove all of this into context. Maybe that is the issue that some people have. Trying to shove everything (prior versions of the same code, etc) into one session.
I always start brand new sessions for every core task or refactoring. "Let's add caching to this class that expires at X interval and is configurable from Y file and dependency injected to the constructor". So perhaps I'm unintentionally optimizing for AI but this fairly easy to do and has probably led to a 5-10x increase in code I'm pushing.
Huge caveat here though, I mostly operate on service/backend/core lib/api code which is far less convoluted than web front-ends.
It's kind of sad that front-end dev will require 100x context tokens due to intermingling of responsibilities, complex frameworks, etc. I don't envy people doing front-end dev work with AI.
As a user, I've found that researching the same topics in OpenAI Deep Research vs Perplexity's Deep Research results in "narrow and deep" vs "shallow and broad".
OpenAI tends to have something like 20 high quality sources selected and goes very deep in the specific topic, producing something like 20-50 pages of research in all areas and adjacent areas. It takes a lot to read but is quite good.
Perplexity tends to hit something like 60 or more sources, goes fairly shallow, answers some questions in general ways but is excellent at giving you the surface area of the problem space and thoughts about where to go deeper if needed.
OpenAI takes a lot longer to complete, perhaps 20x longer. This factors heavily into whether you want a surface-y answer now or a deep answer later.
The shift may be less about fundamentally new content and more about explicit labeling and marketing of elements that were previously present but implied -- i.e. -- SEO.
SEO contributes for sure, but I would reverse the statement here - it's more about new content and less about SEO. There's a feedback loop race to the bottom dynamic regardless
So I'm actually kind of lost here. As I understand it, your theory is that the anticipation of SESTA/FOSTA caused there to be more professional porn and less amateur porn on Pornhub, which in turn caused a "race to the bottom" in porn titles, along both rough/violent/rapey lines and incest lines at the same time, and furthermore that the titles reflected the actual content?
So you believe SESTA/FOSTA led to an unintended increase in the actual violence you'd see if you watched random videos on Pornhub?
I don't think much of SESTA/FOSTA, but I do think that's kind of a stretch.
This means that the person with immunity simply has to keep the level of their crimes low enough that sane and rational people would not jump to vigilante justice. A bad actor could do quite a lot within those boundaries.
The "America, love it or leave it" tactic? It's intellectually dishonest and shortcuts any kind of debate or thoughtful discussion. Or is this more of the "you haven't left your wife beating husband so you must like it" tactic? That's also philosophically bereft and avoids anything substantive.
Suffice it to say, constructive criticism is vital for democratic improvement.
> If your belief is the United States is so bad that it justifies murder, you should leave.
That's a weird conclusion. For me, it's rather "The USA has its flaws (for me - healthcare and higher education financing above all) so we as a society should focus on fixing these problems". Killing people or leaving the country are not solutions, they are are an equivalent of short Twitter replies on a nuanced subject.
Isn’t this how it is supposed to work in America? People own guns to fight tyranny. The gunman carried out his own judgment, but that’s the whole point. And there is reasonable belief that the CEO is responsible for a lot of suffering and expected life lost.
- The intelligent individual is also self-absorbed and believed that they would be able to continue to kill CEOs without getting caught. A narcissistic streak that allowed them to make no attempt at concealing their identity in public. They kept the weapon in order to move to a new target (or they 3D printed an identical if the reports of a 3D printed gun are correct). They believed they would either not get caught or that the public would not turn them in. They may have envisioned themselves the Ted of Healthcare.
reply