HN user
i'd love a series hybrid version of this. to match performance of that I4 you can probably fit both electric motor and gas generator in a normal sized engine bay.
like a compact pickup version of what Edison Motors is doing to the fullsize pickup. i'd take an Edison conversion of an early 90s F250 over any new truck on the market.
Let me know how I can help you all let it rip.
i'd love to see a way (forum, competition, ?) for people to compare harnesses in different domains. folks like accountants, retail store owners, electrical engineers, and all sorts of niches are building personal harnesses/toolkits around claude/codex. those toolkits fit neither their domain communities nor SWE-heavy harness spaces, and a generic home for small niches could help them flourish.
a current problem is nomenclature. i suspect that many people have organically grown toolkits substantial enough to call harnesses, but are not close enough to the SWE bubble to be familiar with the word harness. ai has made it easier than ever to build bridges between domains, but it's also made it easier than ever to get domain tunnel vision and reinvent a super great wheel.
Ender's Game did a pretty good job of answering that question
i've never managed to gather much of a crowd around this soapbox, and i appreciate you standing on it!
i used custom TRUE/FALSE defines that were maximally-orthogonal-without-being-bitshiftable for a while (0b10011010/0b01100101), but eventually gave in and accepted the 0/1 paradigm the rest of the world uses to avoid friction with every library in existence. the only alternative is to become an ada programmer, and my hair isn't quite grey enough for that.
thoughts on bitflip versus bitshift resilience?
stop assigning your self worth to the quality of your work
my personal riff on that is 'stop assigning self worth to the quality that others assign to your work'. being able to internally generate self-worth through the act of producing high quality work is a super power.
those are rookie numbers, my agent harness goes way more than 3 loops deep, and the depth dynamically ebbs and flows throughout the development process. like most brags, i'm being defensive over a personal tool that i know has ballooned way too big, but building a token-fueled katamari ball that rolls itself uphill has been a fun way to imagine Sisyphus happy.
a harness is a chain connecting two other chains, the LLM and the human are themselves made up of loops within loops within loops.
or perhaps infinite jest once it hits tiktok
https://en.wikipedia.org/wiki/Infinite_Jest
a film, Infinite Jest, also called "the Entertainment" or "the samizdat". The film is so compelling that its viewers lose all interest in anything other than repeatedly viewing it, and thus eventually die.
yeah, it was a joke referencing this link that was on the front page a couple days ago,
https://katamari64.se/posts/2026/odin-wikipedia/
a little over the etiquette line for HN, the lack of joke comments here is awesome and i shouldn't be part of lowering our collective brows
a random video that some ginger dude put up on youtube doesn't seem like a reliable source. any info on this from a reputable institution?
shame that HN can't meet the moderation quality bar set by wikipedia
academia will (hopefully) be an industry that benefits from, rather than degraded or obsoleted by, ai. the system should self-correct. scientists that use ai to give themselves more time to think will succeed and scientists that use ai to avoid thinking will fail. over the next few years processes and pedagogies should develop that help reliably guide researchers toward success.
there's a lot of tedium in academia. ai taking menial tasks off the plates of grad students and post docs allows them to spend more time collaborating with, instead of serving, the head of the lab. dissertation defense is already a strong mitigation against producing folks like that satirical Dr. Simmons, you can't defend a dissertation if you don't have the chops for a chalk talk.
anyone have a good argument that steel-mans the anamnesis blog being ai slop?
an interesting coincidence is that the blog's july 3 post ties into the wikipedia/odin kerfuffle that's on the front page. its unorthodox take on doubting Thomas suggests we believe "a trustworthy witness without demanding to rerun his experiment yourself". sounds like the blog sides against Jimmy Wales that in our world of agenda-driven primary sources, trustworthy first-hand accounts are the best sources we have. it goes on to doubt all sources other than Him, including itself and yourself, and that's probably not an actionable moderation policy change for wikipedia.
https://katamari64.se/posts/2026/odin-wikipedia/
https://anamnesis.blog/posts/2026-07-03-reach-hither-thy-fin...
you're correct that being good with language is a useful skill for operating LLMs, and fortunately you get to practice and improve that skill whenever you work with an LLM. being well-read helps too, being precise with vocabulary is important and idioms/metaphors can pack a lot of semantic meaning into a few sentences.
just because you don't know how to swim doesn't mean people can't swim.
$250 doesn't buy you out of them monetizing your data. last year they started selling what content is being watched. next money grab will be selling the list of what's in your media library, all they have to do is update their terms to allow it, like they did last year.
agentic coding is a soft drug, so taking 'the only way out is through' approach is pretty viable. once you figure out how to swim and have claude running for an hour+ at a time and only bugging you with either high-level taste decisions or 'done, how's it look?' it's pretty low stress
if you're overloaded with PRs, build LLM-based systems to take the load off. don't be a senior engineer, be an engineering manager.
isn't this the expected/intended outcome of the economic policy we've held for a generation or two?
the capital/labor class gap increases when total returns on capital investment exceed total wages+redistribution to the labor class, and the gap shrinks when that's reversed. the market ~controls the capital gains and labor wages knobs, and society decides where to set the redistribution knob.
the article investigating the post-covid drop and concluding that it's normal is an interesting rhetorical device. on one hand, relief that nothing crazy is happening. on the other hand, disappointment that we've accepted a growing inequality gap as normal. the gap was already at a post-war max going into covid, the floor gave out 20 years earlier and covid was just gas on the fire.
advancements in automation and tax codes that benefit capex over payroll will continue to incentivize business to shift budgets from labor to robots.
OpenAI/et al. selling an IP laundering service under the name 'max subscription' may force the world to accept the perspective that Intellectual Property isn't a thing. The business model of extracting value from creators via rent seeking IP may not be viable in a world where LLMs can generate anything on demand. We might be transitioning to the Lockean view that for something to be ownable as property, it must be a scarce resource, and information is not a scarce resource.
From that property rights perspective, the property that's created when new information is created is not the information itself, rather, it's the act of creation (claim to authorship) that's the scarce resource.
I don't know what a world looks like where the only form of IP is non-transferable and owned by the original creator. Maybe that new form of IP creates less value over all, and maybe that's ok if the creator is getting 100% of the smaller pie instead of crumbs from media labels. Companies like Red Hat could be an example of a viable business model if IP laws follow the current winds.
Companies like Corgi will need to rely on internal talent to ensure that their product is better than what someone looking at their product can vibe code a copy of, which from my perspective as a consumer, sounds like a better route than Corgi relying on an internal legal team to send a cease and desist letter.
the same technique you use to get a human engineering team to generate a deterministic and repeatable result 100% percent of the time. i know that's a flippant (but true!) answer, it's just a pretty tough problem :) setting the bar lower than 100% makes it a lot more tractable
for me, it's less 'still' and more 'again'. claude code + API tokens means i no longer have to suffer the user-hostile design of many webpages. using full-screen claude code feels like finding my old DOS teddy from childhood buried in the back of a closet.
this is the type of thing you need to build a foundation sturdy enough to let you operate higher up the stack and ratchet to design-by-metaphor and then design-by-philosophy. those design skills are taught in humanities departments, not engineering departments, so this is a weird feeling place for those of us that wandered over from a technical field.
yes, that is the messaging that meta leadership is going for, you have received it as they intended
labor shortage and high unemployment can, and will, coexist. the roles most in demand will be those that fewest of the population are able to fill.
because they need to make as many employees quit as quickly as possible if they hope to avoid bankruptcy. i don't think they'll be able to avoid that outcome, instead they'll die trying. maybe their capital and network will be enough to buy them time to fully pivot to hw, though that's probably less of a moat than i believe it is while wearing the rosy glasses of an EE
software is now free, at least for the people that know the proper incantations required to manifest it into existence. software-only companies have no future. sending old-world SWEs into the undiscovered country results in high costs to the unprepared SWEs and high costs to their financiers who lose control as soon as the boots on the ground realize the wildfire is too close for comfort and new winds are blowing
the only viable way to separate assets from liabilities (payroll) fast enough for large corps to catch up with the growing number of claude-unicorn centaurs, and small herds of them, will be bankruptcy (could be wrong, i'm no lawyer)
it sounds like we share a vice for information gluttony. i have a similar workflow where i continuously scan the digital horizon for useful ideas and continuously feed them into my personal toolkit. i've kind of always been like that as a person and an engineer (the two aren't really separable for me), and claude code has been like switching from coffee to meth, and, um, there's some downsides to doing meth. it's hard not to build a house of cards that you end up abandoning when it starts feeling wobbly and you see a greenfield next door
no one knows what that link looks like but you
as an example, i built tooling that lets me open claude code and say 'add MSP430FR2476TRHBR to the library' and a few minutes later there's a schematic symbol and footprint in my kicad library with parameter fields consistent with my organization's internal standards. i'll be waiting a while for anthropic to add that feature to claude code.
agents are everywhere nowadays, one left a long pointless comment on a bug report i submitted on github. well, a bug report that an agent submitted on my behalf. agents all the way down. maybe i'm part of the problem.
This reminds me so much of my own experience with AI fueled dev mania. Rapidly build semi-functional wonders, then pivot to something shiny and new to avoid the QA trudge of polishing that wonderous turd.
I'm on my second agent orchestration framework, Omnispect - https://omnispect.dev/
Example created by Omnispect:
Oneshot - https://omnispect.dev/battleclone00.html
Polished - https://omnispect.dev/battleclone04.html
I'm impressed with Opus 4.5. It's been useful working on firmware projects where earlier models were of negative value.
Here's an example of a one-shot output, the only change I made was Replace All 'battlezone'->'battleclone':
"build a clone of the classic arcade game battlezone using SVG graphics that are calculated on the fly for the required vector wireframe graphics"
It uses your existing subscriptions and supports all the major CLI and API providers. There's no cloud features of Omnispect itself, it runs locally except for calls to the LLM providers.
Claude Opus 4.5 is used as a routing agent, which selects the most appropriate LLM provider and model tier to delegate a task to. For example, the routing agent might delegate a single large task to GPT-5, which in turn delegates multiple small tasks to Haiku agents in parallel, then Gemini reviews all the work.
Omnispect lets you view the delegation tree of prompts and responses that spawn from your initial prompt.
A desktop application to visualize and document how tasks flow through a dynamic heirarchy of AI agents.
This is really cool! Development environments that facilitate low-friction use of many models and providers are awesome.
Here's something similar that I've been working on - https://omnispect.dev/