Anomit

Generative LLMs, by nature of their architecture, won't one-shot or n-shot (n < 5 (anecdotally)) completely bug free code generation. Asking another LLM to generally "code review" is a probabilistic gambit in itself. Definitely requires burning tons of tokens on frontier models to get anything useful out of the effort. Such efforts are to be expended to what end when we can get to a clean operational code base with deterministic guard rails? The issue is intellectual laziness. System architecture and production scale engineering, security etc patterns and lessons are being forgotten or most have never even studied in the last couple of years. "make no mistakes" does NOT cut it. The other path? Specialized fine-tuning of open weight models on synthetic data and fixtures so the model makes fewer mistakes up front. That will still cost more compute than just forcing the codebase through deterministic checks. Human eyes and judgment were the deterministic factor earlier. The velocity and volume of code generation with LLMs needs a whole different approach in the present age.

Yuuuge pile of changelog shipped @CleanSlop in the last two weeks. Went live on Product Hunt with a fully agentic workflow. Backed by Stripe Link agent identity. 95% faster scans (yep) Autoscans are live. And take a look under the hood of the system at cleanslop.com/under-the-hood

Didn't these folks get hacked and went down for more than a day? And now they are gaslightmaxxing? Bring back some goddamn shame ffs

Got my hands on a console (DDJ-FLX4) for @divyaranjan_ 's house party. Couldn't play that night because of a USB drive failure but the itch was too strong to ignore. Mixed a set after 3 years once I got home. Glitched twice in there, but I still got it. mixcloud.com/0xSha/bad-groov…

proliferation of matcha pilates padel runclubs is just the sort of indicator that precedes the crystallization of such a tone policing society. I mean you are already getting lectured on tone appropriateness when having to interact with misanthropic models. Only a matter of time

Remember that time when blockchains wasted a good half a decade moving to PoS from PoW because : mUH enERgy conShUmptIOn :`~/ ?? Now we got hyperscalers, datacenter spanning multiple towns and suddenly energy efficiency and ESG scores are forgotten. Everything is such a psyop

Built out this local tutoring webapp for a friend for CISSP studies and cert exam. The challenge is to have continual testing along with lessons that she can follow while juggling challenges with full time work and family commitments. If it works well, I will open source it.

The heady mix of exhaustion and accomplishment is real today. The second major, 100% agentic, release of @CleanSlop is on Product Hunt today. It supports Stripe's @Link agent identity to attest to the validity of an agent. We are among the first few products at the moment to integrate it throughout our workflow. And cleanslop.com is a first-class agentic resource at this point. Just point your agent or assistant to it in the context of the codebase you want to work on. Also, definitely a good time to clear out recurring queries and refrains that we keep running into. (blogs and docs will be up once the dust settles from handling the influx from this release) - This is a deterministic risk scanning and surfacing engine that has been trained on 25+ categories of where and when things can go wrong. How do we know where to look? Just check the scale of things we have worked on in the last decade. - Techincally, it is static analysis. We haven't gotten to the runtime yet. And it is currently monofocussed on vibecoded webapps that are typically generated on Next.js/TS/Supabase sort of stack. iOS support coming soon. - The engine is continually trained, updated and refined on widely available public corpora of codebases as well as synthetically generated ones. We can promise our results and findings will only get better with time, not regress. - WE ARE NOT A LLM WRAPPER AND YOUR CODE DOES NOT GET SHIPPED TO RANDOM LLM PROVIDERS. Can not stress this enough. We have a ZDR in place if you don't want to include your codebase for training purposes, and when it comes to multi-phased scans where the scanner surfaces findings it is not confident on, only slices of callgraphs, import trees etc are consulted with decision models like @typesafeai Jev or @Cloudflare Clef. - This is NOT a security audit. This is not a package or collection of SKILL.md files that require an LLM subscription from you. This is definitely something that will bring to attention what will break when internet scale traffic and users hit your app.

Decentralization was supposed to solve the fiascos that have come out of centralized dependencies. Filesystems, package repos, mirrors, name roots (I recall Handshake Protocol). Instead we let zyn coke addicted Wall St. quant imports to ruin the show with their casino greed BS.

Google has refused to acknowledge issues with their Pixel 8 Pro WiFi for almost a year. And I continue seeing major system updates pushed at least once in a month, sometimes twice. At this point, it has become an unusable brick. Planned obsolence is a real thing.

Working with @SwaroopH is super fulfilling on this. His crazy product brain took over the harder work of exposing the @CleanSlop backend harness to multiple carefully crafted pathways of integration. A swiss army knife agent CLI, login via Link or GitHub, MCP, local code uploads, 402 standard payments via Link. And there is more on the way. Watch out for next week's shipping updates.

I got early access right the night before we launched @CleanSlop on Product Hunt. "if you decompose enough, model providers can't even figure out what you're working on" Speaking of decomposing queries, right after the launch, the first thing I did was get on to was running a complete audit on our scanner engine with Jev and finding out the leaky reasoning pathways. nitter.cf/anomit/status/21009493…