I suspect that a lot about what we call being very intelligent is ultimately out-remembering people around us. I think of all the times in my software career when I did something that others considered very high performance, it either came down to either having more energy than others at tackling a problem they thought was more trouble than it was worth, or just bringing back random knowledge from previous jobs or self study, and being able to apply it to the problem at hand.
I don't think I've had a truly original idea in my life. Combine A + B, when it's rare for people to know A and B at the same time. So from that perspective, what LLMs are doing is basically the same thing. Sometimes I am faster than the LLM because my context might be better organized, but it typically needs just a hint from me to steer itself correctly. It claims something is a memory leak, but smelling a rat, I suggest it to double check the garbage collection statistics too, at which point it's clear it's no leak, but a tuning error, at which point the LLM is better at tuning than me, because it has more energy than I do.
Maybe there's true brilliance out there, when something doesn't come out of combining data and building hypothesis until you get really lucky. My experience is not comprehensive. But I look around me, and it sure seems I've not been lucky enough to see it. Even the shiniest people I've worked with, which most of the audience here would recognize, have never shown me that they can go past this.
tdlr: This is a Novo Nordisk-funded study focusing on predictive biomarkers rather than real-world dementia cases. Novo Nordisk's actual dedicated clinical trials for Alzheimer's completely failed to show that semaglutide stops cognitive decline.
"A predictive biomarker is like a "check engine" light on your dashboard. It warns you that there is a risk of a future problem. In this study, the researchers only checked if the drug turned off the "check engine" light (by measuring blood proteins), rather than testing if the car was actually driving properly (by testing the patients' actual memory and brain function)."
Always do FIRST analysis on studies. Or have AI do it for you. I used Gemini to dig into this:
"Novo Nordisk funded this study, and several of the researchers are employees or minor shareholders. While corporate funding doesn't automatically mean the data is fabricated, it does mean the company is highly motivated to find and publish data that makes their blockbuster drug (semaglutide, marketed as Wegovy, Ozempic, and Rybelsus) look like a preventative treatment for a wider range of conditions, expanding its market and driving up profits."
"Funding: The study was funded by Novo Nordisk A/S.
Investigation: Researchers conducted a post hoc analysis using data from the randomized, placebo-controlled SELECT trial. They applied the Dementia SomaSignal Test (dSST)—a 25-protein risk score—to non-fasted serum samples collected at baseline and at week 104 to estimate 5-year and 20-year all-cause dementia risk in patients receiving semaglutide (2.4 mg) versus a placebo.
Results: Semaglutide significantly attenuated the progression of the dementia risk signature. Compared to the placebo group, the 5-year predicted risk increased 2.5-fold less (a 26.0% lower predicted event rate) and the 20-year risk increased 1.67-fold less (an 8.8% lower rate). Semaglutide also lowered the odds of patients moving into a higher dementia risk category by 36%.
Subjects: The analysis included 2,970 older adults aged 65 and older (mean age of ~69.7 years) who had overweight or obesity and cardiovascular disease, but no history of diabetes. The cohort consisted of 814 women (27.4%) and 2,156 men (72.6%).
Time: The study evaluated data over a 104-week (2-year) follow-up period. The analysis was published on August 8, 2026."
And then map the weakness to each respective letter if you want to dig deeper.
The word is "management", not "leadership". This comes across as a LinkedIn post filled with vague notions and weak writing.
The conclusion also completely contradicts a previous point, which is that managing an LLM is not like managing a human. So the skills are, in contradiction to that LLM-ism of a conclusion, new. The author isn't using their people management skills, they're using new LLM-management skills. They think the two are similar, but didn't bother breaking down how they're the same vs where they contrast. It's just a lazy observation expanded out to a short essay that says nothing interesting.
People have told me I was smart since I was a kid, but I can't remember for shit. I had a thought when I was fairly young that the only reason I was (maybe, sometimes) outperforming others intellectually is that I was habitually compensating for my poor memory by working things out on the fly, while others could rely more on rote memorization. Anyway, takes all kinds I guess!
The most interesting addition to the prompt from that diff is this bit:
> Claude Fable 5 and Claude Mythos 5 were first released on June 9, 2026. On June 12, 2026, Anthropic suspended access to both models to comply with U.S. Department of Commerce export controls; the Department lifted those controls on June 30, 2026, and Anthropic restored access on July 1, 2026 (Anthropic's statement: [https://www.anthropic.com/news/fable-mythos-access](https://www.anthropic.com/news/fable-mythos-access)). These events are after Claude's training-data cutoff, so Claude knows about them only from this notice. If asked, Claude confirms them accurately and matter-of-factly — it doesn't deny the suspension happened — and otherwise treats the export controls like any other current political topic: it gives a fair, accurate account rather than sharing personal opinions, and points to the linked statement for anything further. Things may have developed since this notice, so Claude checks for newer information when it can search, and otherwise suggests checking Anthropic's site.
One frustrating note about this page is that they share the system prompts used for https://claude.ai and the Claude mobile apps regular chat, but they omit the tool definitions. Those are much more interesting if you want to understand what Claude can actually do for you. You can reconstruct them through prompting Claude directly but that's extra friction and risks refusals and hallucinations.
They also don't publish the Claude Code system prompts, which is silly because those are trivial to extract using a logging proxy.
In the last couple of days I wanted to try out the new definitive DeepSeek v4 releases. I gave it the repository of a semi-abandoned video compression codec and I told it to perform the usual benchmark -> profile -> verify -> research -> improve loop. I specifically chose this codec because the authors include a verifier for the bitstream to make sure you don't break stuff if you want to try your own implementation. I gave the agents access to the compiler's profiler and also Intel's VTune, which has fantastic output. In a couple of hours the LLM generated SSE and AVX implementations of the compression and decompression algorithms that almost doubled performance with a single core. Then I asked it to create a CUDA implementation using NVIDIA's NSIGHT profiler as a guide and it also started doing some good work.
Personally, I believe that LLMs should be treated like an advanced version of Prolog or linear programming: you give the constraints, you have a way of verifying correctness, and you give it a clear goal. If the LLM can verify itself and course-correct you can basically leave it on autopilot
Offtopic. I have a concern that this forum is removing stories that have negative connotation on AI.
Few days back, I posted an article[1] that was about how AI threatens natural resources for billions. This was from United Nations and it was flagged. I did not think much about it until I saw two other stories [2] & [3] today that were doing fairly good on front page but they suddenly disappeared. They are not even on 2nd or 3rd page. I have seen this happening at other times as well but did not document it. Just thought you all should know about this.
I was going to create Tell HN thread but I thought the same would happen with it too. I am pretty sure this thread is not going anywhere so I'm posting my concern here.
<quote>
Have you ever heard of the Joined Together States? Or bosom peril? Kidney disappointment? Fake neural organizations? Lactose bigotry? These nonsensical, and sometimes amusing, word sequences are among thousands of “tortured phrases” that sleuths have found littered throughout reputable scientific journals.
They typically result from using paraphrasing tools to evade plagiarism-detection software when stealing someone else’s text. The phrases above are real examples of bungled synonyms for the United States, breast cancer, kidney failure, artificial neural networks, and lactose intolerance, respectively.
</quote>
It's also "out-brute forcing them." It just never gets tired. If a mathematician picks a research direction and spends a whole week on it and it doesn't pan out, they will likely be annoyed, need a break for a while, etc. This thing just does not ever get tired or discouraged or care; it's just onto the next thing until something ends up working.
Ideally what I'd like to see is pluggable knowledge bases.
So if I'm e.g. coding a SwiftUI app for navigation, I'd take 9B of basic coding and reasoning, add 10B of swift/swiftUI, add 5B of GIS/geography knowledge and another 5B of frontend app design knowledge. My model doesn't need to know a single line of python.
Then when I want to research electronics components, I grab a 15B model of agentic research techniques, and add in 10B of electronics knowledge, etc.
I don't want general purpose models. They try to be everything to everyone. I want to click together a model that is laser-focused on what I am doing, and I want to run it locally
Please understand that the goal of these policies is to weaken scientific research in the US. The people who push this stuff acknowledge openly that they oppose science, experts, and accurate information. This isn't a misunderstanding or a fumble.
I think he's kind of speaking past the original author. The original piece is basically about how the author doesn't think that RISC-V will take off outside embedded, because of some design decisions that lead to poor performance compared to ARM64 and because so much of the ISA being optional means that there's too much fragmentation to make binary distribution feasible. Meanwhile, this piece is mainly about how RISC-V is great for embedded because companies can build it into custom chips with specifically the functionality they need, and because of how cheap it is for low-end use cases since there's no license fees.
The only real point of contention I see between the two is that this piece goes on to talk about how it's a selling point that RISC-V can be used for both low-end 10 cent microcontrollers, and high-end multi-core processors running Linux. Personally I don't see the benefit of this since you're going to have to recompile your software anyway, and since all the RISC-V SBCs I'm aware of have significantly worse performance and efficiency than comparably priced ARM SBCs.
An LLM works better the more disparate world knowledge it has, even if it's not immediately obvious why it would be relevant. The model finds a structure to the problem you give it in a largely language-agnostic way that benefits from training on every language (these things are direct descendants of Google Translate), and even non-programming knowledge - the structure of your task might resemble an ancient Chinese poem that influences the model's response, for example. That structure is considered a form of compression, as some fascinating and illuminating recent 3blue1brown videos get into - a common pattern in Haskell or FORTRAN and a situation described in an ancient Chinese poem may all compress to something quite similar to your task, thus when the model compresses the idea of your task it immediately draws from those ideas.
There are "experts" which do divide parts of the model that are found to activate together for specific tasks, so they can be processed in parallel to join the result at the end, but it's nowhere near the granularity of a SwiftUI expert and a python expert. The difference in those things is so trivial from an abstract point of view that it would make no sense. They would be 99% the same.
Distillations also come into this but I'm highly skeptical you could make one guaranteed to only know programming and only in one programming language (especially with as small a sample set as SwiftUI relative to something like C) without its efficacy being hobbled by tunnel vision. Reminiscent of the SpongeBob episode where he empties his mind of everything except fine dining and breathing, then can't remember his name and goes insane. Beyond the basic concepts of general coding and the trivia of syntax, getting anything done requires a large intersection of disparate world knowledge and the ability to apply it to new situations.
> I want any LLM I use to choose the very best, most precise words at every single decision point.
Then bad news: LLMs already use randomness in a fundamental way. Each time they go to generate a token, they first generate a probability distribution of possible tokens. Then they pick one randomly according to this distribution. The technique described can be thought of as making the random number generator pseudo random. The output it generates is one of the possible outputs it would have generated before, just now it's deterministic and will generate the same thing every time.
Dropped rods are an incident but one that occurs because of pressurized water reactors being very default safe. Controls rods are one way the criticality of a reactor is controlled and US reactors (in general) will go sub critical if even one rod is fully inserted into the core. You will likely have heard of a reactor scram (which goes back to the safety control rod ax man) where in an emergency, all the rods are dropped back into the core, greatly reducing its criticality. In some cases, an interruption of electrical power will cause a rod (or three) to drop accidentally. This is a "dropped rod" incident and will force a reactor shut down because it is now sub critical.
Lots of knock on effects -- sub critical, let heat in the primary loop, less steam and electricity generated in the secondary loop, etc -- but generally a non event that you practice for.
There's no reason this would lead to a radiological event or more significant casualty.
The problem with this analogy (actually one of many) is that most people can make do with a cabinet that is literally identical to everyone else's. IKEA is great for that.
If you want software that is literally identical to what someone else is using then you don't need AI. You need a license to that software! That is just the traditional software model.
AI gives software that is bespoke with hundreds of decisions made, hidden from you, in the background. If it's a throwaway script, that's fine (and I don't mean to undersell this - this is a huge application). If you want a larger program that is going to form part of your business process then it will need at least some level of supervision from an actual expert.
It'd be ironic if the "Opus 5 nerf" effect is from telling Opus that it sits a tier down from Fable and Mythos, while Opus4.8 believed it was the best of the best, just a note that it was "Preceded by Mythos".
PgBouncer is entirely optional and it's not always the right choice. If you have a classical app (non serverless) and you can maintain a connection pool from your app, then I recommend avoiding pgbouncer.
The benefits of pgbouncer mostly come from irregular client connections (too many, too much churn). If you don't have that problem, go direct to postgres.
I'm exploring replacing pgbouncer with an alternative (maybe home grown) at the moment. Mostly for multi-tenancy and HA reasons. Pgbouncer has been good for us, but it's limited in how we can deploy it in a multi-tenant environment.
My Eng lead has no coding experience, 25 years of management experience, yet has driven 3 separate projects into technical bankruptcy to date.
He just accepts anything that Claude says as truth. He vibecoded over 60,000 lines of code in 3 weeks, but couldn’t get it to do what he want and made a project overrun for 3 extra months. When the pissed off stakeholders called a meeting to ask what was going on he didn’t show up and sent his junior engineer to answer questions and take the blame. Now thats leadership.
Crazy how a smart person like this fails to understand the gumbel softmax technique. It does not affect writing quality at all, provably. The very fact that there is generally no "best next token" with 100% certainty is precisely why the trick works (you cannot watermark a response to "respond with the To be or not to be soliloquy from the first folio Hamlet", for precisely this reason).
To people asking why, this is a good lesson on the Collison’s ambitions. Stripe is one of the best API companies in the world. They know how to serve high volumes of latency and availability sensitive requests. They’ve abstracted the financial rails for payments and now want to abstract the rails for LLMs.
They’re the perfect company to own OpenRouter.
Tokens are simply a lightweight valuable asset. Stripe can serve as the middleman as well as anyone. They know how to route to many providers (payment rails) with huge differences in service characteristics. LLM providers are far easier.
Then they can work this into an offering where users can subscribe to tokens and use them across services. It solves one of the core monetization challenges of every AI company: how do you price when your costs are variable on usage, but nobody can make sense of charging by token?
From here, they can start hosting their own models and competing as an AWS for tokens. They can be the best provider of $OPEN_MODEL, or their own, and optimize for you.
Wow, $2k? You could get a decent ebike for that money. And the installation doesn't look any easier than a hub motor... Only seems worth it if you already have an extremely nice bike. I think of friction drives as being aimed at budget-conscious commuters, not serious mountain bikers.
IKEA furniture embodies many elements of good cabinet making but skips many nonessential elements. And does this more consistently than cabinet makers who can be bored, incompetent, depressed, burnt out, resentful, tired, having a bad day.
In the future AI code inevitably will embody most good software engineering practices. And will do this more consistently than software engineers who can be bored, incompetent, depressed, burnt out, resentful, tired, having a bad day.
Just look at the messages on HN or around you at your colleagues to see how mediocre the average software engineer is..
Today's IKEA is good enough for most people.
Tomorrow's AI coding will be good enough for most corporations.
Good enough to vastly reduce the need for fine craftsmen and women / software engineers.
Good enough to deskill those who call themselves cabinet makers / senior software engineers. These days the cabinet makers I personally know just do contract kitchens for project builders.
But IKEA is and AI will be, bad enough that at the high end with special requirements / taste / money / an inflated sense of self worth, some furniture makers still exist and thrive.
Perhaps 1% percent of current software engineers of today will be needed in the future when AI code inevitably has the ability to follow good software engineering practice.......
And as usual it will mainly be the mediocrities that remain ( so there is hope for you too ), with occasional islands of excellence.
> how AI threatens natural resources for billions.
The article rests on the claim that water usage of data centers on continent X threaten water availability for humans on continent Y.
I hope you can see how self-evidently illogical that is. The article tries to bend logic into a narrative it is trying to push.
> This was from United Nations and it was flagged.
Indeed! It exposes a level of lack of rigor and critical thinking which is astounding - this is meant to be an organization that thinks clearly, which it clearly does not.
“Subsequently, 1 mL of the mass killing of an ethnic group was opposed to 20 mL of the skin sample and unprotected to light for 7 min.”
From: https://bsky.app/profile/forbetterscience.bsky.social/post/3...