joshuamoyers

so ironic to see “load bearing” in a comment about removing AI. certain phrases have been ruined by fable/opus 5 for all time.

i think 1 billion DAU is the exception here, so I would not expect everyone to constantly caveat personally.

YMMV but I think its going to always be important to acquire almost all the knowledge necessary to go to the metal, then allow the tactical to bits to atrophy over time. What remains is all the bones of the first principles understanding - and that will serve you well indefinitely.

An analog for you: I started in software pre-cloud, with a primary focus on web technologies, but at various times learned c/c++/asm/compilers/operating systems/algorithms. I audited many classes, but one instructive one was a course called nand2tetris (going from nand gates to hardware simulated CPU to peripherals to operating system to programming languages and compilers to application layer). Understanding the global through lines in computer science give you immense power.

Does my work as a CTO mean I use this information all the time? Absolutely not. Did I benefit from figuring out how computers work generally throughout literally my entire career? Absolutely yes.

I am actually relatively AI-pilled compared to the average Hacker News reader, but I am against outsourcing fundamental understanding to AI in all forms. Does that mean I write code by hand often anymore? Very infrequently.

I think it'd be very hard to achieve viable tokens/s or get arithmetic intensity to be high enough in general, since many cases in existing training and inference are memory bandwidth limited. Definitely seems possible to conceptually have a slow pipeline that is distributed though.

it seems to be something that’s similar to the class of optimizations associated with with linear or state space attention when things models often do is once they figure out an optimization like this they create a ratio between full resolution blocks and blocks that have the optimization implemented.

I think the primary reason it works is because the difference between K and Q, which is not all that obvious is that it’s allowing the model to have an asymmetric relationship between tokens, so one token can attend to another without the reverse being true. It seems to me if you just have a single value that you’re representing symmetric relationship, which might degrade the quality of reasoning over a set of tokens, but also is probably possible.

article points out a real problem - simplicity is one of the hardest things to achieve. the act of reduction is important.

buts its a refreshing that there is an initial list of half baked projects, i suppose meant to evoke horror at the untidiness and wasted time. but honestly each of those projects sound cool as hell. not necessarily durable - but who cares. i’d argue there is a skill, one that is different than traditional programming, that the author was building up over that period.

discipline is important. focus is hard. but allowing yourself to play is not a bad thing at all and i dont think building little interesting side projects should be a shameful act.

we’ve been struggling with review throughput. this actually seems worthwhile to build at this point though i remain fairly skeptical of workflows that are agent-only, at a point it seems like the only practical solution.

we are finding lots of value in self review. its the “imagine you are doing a synchronous paired review with someone - anything that is difficult to explain, has a code smell, doesnt fit the architecture of the system around you, write a comment.” then at the end, agents do a good job of looping over PR comments.

the second thing would be a guided, educational code review tool - there are a few attempts at this, but nothing that has a good enough interface to actually stick. organize hunks by semantic importance, spend some tokens exploring the surrounding systems, showing how new code, public apis and data model flow with the existing design, and allow a human to traverse larger PRs more quickly.

thank you to cloudflare for publishing this.

> But for me, it becomes the event my whole day starts to revolve around. I have to break out of my flow, put my tasks on hold, take the call, and then get back into context. In the end, a 10-minute call can cost me several hours of focus.

Occasionally I get this feeling for a large customer meeting or a public talk, because there are consequences and serious prep. But this is just trying to normalize extreme social anxiety and call it a management style.

One reason you get together to talk is so you can hash out details on potentially ambiguous topics, so you don't head in the wrong direction causing net negative contribution.

Another is that people are not automata. Humans require inspiration and motivation and you need to reinforce the vision of what you are building and why. Its also even sometimes a reasonable idea to ask about how their life is going and check up on their family and pets and career aspirations.

In general, some people should not be managers, and there is plenty of room in the world for super ICs.

except agentic engineering mostly invalidates this with regard to marketing websites. nothing is really all that hard to implement on mostly static site.

i appreciate the sentiment to a certain extent - its not going away, skate to where the puck is if you care to do so. but the writing is repetitive and theres an entire repeated paragraph (bullet to paragraph form). there are also lots of things to be worried about even for the most seasoned individuals in terms of half decade increments conservatively. assuming large parts of swe become commoditized in the form of paying a handful of frontier model providers more and more of the share of what was once swe wages, the high end is what survives. high context fox-like (a la terrence tao's foxes and hedgehogs) are guiding ai to build - and then they are eventually displaced as well. extreme societal pain seems like its on the horizon assuming we dont have some incredibly unlikely massive mobilization towards post-work post-scarcity thinking with active social safety nets. economic diffusion probably means this is a little further away than we think, but time moves pretty damn fast.

i love how half the comments are literally doubling down and simultaneously angrily complaining about auto-scroll. hacker news has become worse than mid 2000s irc.

you cringe while simultaneously posting a github link with your “current setup” - do you see the irony?

I think this is the right take. I usually am aligned with most of what Anthropic is doing, but cutting off OAuth login from open harnesses was a bad move. My guess is there is some serious worry/overlap with the Cursor's of the world - e.g. folks who will be competitors in the future who are taking advantage of cheaper Opus rates/loss leader from them while simultaneously building a competitive model (Composer).

Also, nice clever optimization here. Lots of low hanging fruit in harness land.

I really like this approach. Nice job!

> We also plan to compile solved steps into micro‑policies. If you're running something like a RPA task or similar workflow as before, you can simply run the execution locally (with archon-mini running locally) and not have to worry about the planning. Over time, the planner is a background teacher, not a crutch.

Conceptually, I really like this - why re-do the work of reasoning about an already solved task? Just do it again. For some plausibly large majority of things, this could speed things up considerably.

> In the future we hope to run a streaming capture pipeline similar to Gemma 3. Consuming frames at 20–30 fps, emitting actions at 5–10 Hz, and verifying state on each commit.

I love targets like this. It makes you tune the architecture and abstractions to push the boundary of whats possible with a traditional agent loop.

The salience heat map compression is a great idea. I think you could take this a step further and tune a model so that it compresses an image into a textual semantic/interactive element hierarchy. This is effectively what browser-use is doing, just using javascript instead of a vision model.

This seems like a task that would benefit from narrow focus. I'm aware of the "Bitter Lesson," but my intuition seems to tell me that chaining together fit to purpose classification as an input to an intelligent planning system is the way to go.

This was pretty wild. Veers deeply into broad generalizations that have the potential to be dangerous in some way I cannot name - but if you stop and consider each archetype as a vector along which you can accidentally trap yourself, its a thought provoking read at least. Unfortunately its also a list of undesirable damaged characters followed by some model of a "whole man" that is somehow infinitely attractive and stable. That's a lot of malarkey in my opinion. We're on a many-dimensional journey and all of us are some degree of lost. Some good guide markers in here though.

I think it does boil down to "try things a lot," especially creating real connections with other people, even though you will painfully fail many times. Drive yourself to have real conversations. Protect your health and keep yourself strong physically and mentally. That's a powerful base to be standing on. Then go find a blend of interest, purpose and duty - building a sense of dharma helps you wake up in the morning and move through the world feeling a little less "lost."

> For all that, GPT-5 is not a terrible model. I played with it for about an hour, and it actually got several of my initial queries right (some initial problems with counting “r’s in blueberries had already been corrected, for example). It only fell apart altogether when I experimented with images.

Spatial reasoning and world model is one aspect. Posting bicycle part memes does not a bad model make. The reality is its cheaper than Sonnet and maybe around as good at Opus at a decent number of tasks.

> And, crucially, the failure to generalize adequately outside distribution tells us why all the dozens of shots on goal at building “GPT-5 level models” keep missing their target. It’s not an accident. That failing is principled.

This keeps happening recently. So many people want to take a biblically black and white take on whether LLMs can get to human level intelligence. See recent interview with Yann LeCun (Meta Chief AI Scientist): https://www.youtube.com/watch?v=4__gg83s_Do

Nobody has any fucking idea. It might be a hybrid or a different architecture than current transformers, but with the rate of progress just within this field, there is absolutely no way you can make a prediction that scaling laws won't just let LLMs outpace the negative hot takes.

100% agree. I feel like this is a symptom of Dead Internet Theory as well - as a negative take starts to spiral out of control, we start to get an absolute deluge of a repurposing of the directionally negative sound bytes and it honestly feels like bot canvasing.

> By being particularly bad at anything outside of the most popular languages and frameworks, LLMs force you to pick a very mainstream stack if you want to be efficient.

Almost like hiring and scaling a team? There are also benchmarks that specifically measure this, and its in theory a very temporary problem (Aider Polyglot Benchmark is one such).

> companies often over-index on crystallized knowledge over fluid intelligence.

another way to say this: focus on aptitude. in my hiring funnel, this is a core tenet. you need to be able to capture polyglots and systems thinkers. its still pretty hard to design a process that balances this all very well. combine that with an absolute glut of applicants and you have a very challenging problem.

Its not at all imo. Franchised businesses are not in the habit of employing low skill workers as a public service. This data is interacting with both covid effects and infrastructure upgrade/rollover - in other words, it takes a while for companies to adopt affordable touch screen ordering systems and its been phased in at a ton of non-fast food (at least in my area) over the same period of time. Local health grocery store has touch screen ordering at their deli, as well as simultaneously going cashless. Most coffee shops too. Look at most international airports - almost all the kiosks have one or no attendants now.

Meanwhile, current administration gutting the clean air act: https://www.npr.org/2025/07/29/nx-s1-5463771/epa-greenhouse-...

two good points there are very intuitive - a fresh perspective yields better results and once you are stuck (e.g. 80 iterations) its better to just start fresh. i've seen the same thing anecdotally in coding sessions where context needs to be compacted multiple times. its usually just better to start a fresh conversation and re-seed the basics in the conversation.

In light of recent light controversy around Andrew Ng's talk, I thought I'd submit this long form article I wrote over the last few weeks. By contrast, its my very tactical and specific experience using agentic approach to high throughput software development in a >1M LOC codebase. I've recently also been doing roughly half my work with Claude Code vs. Cursor. I do think Andrew Ng's take on product management (and also design in our case) being the new bottleneck at current ratios is correct. I'm not sure what the best solution is to that in our team's case, but we are definitely feeling it.