Greg Val

@om_patel5 this is the failure mode nobody priced into their stack assumptions. Companies treat AI accounts as critical infrastructure, but the contractual setup is closer to a loyalty program the vendor can pull at any time

vendors get to design the gap between announcing an infrastructure decision and the day it costs you

@NabilChiheb the silence at month 3 is the dev recognizing there's no architecture to read, just a sequence of patches that each made sense in isolation when the ai wrote them. haven't cracked how to keep an architecture readable while ai writes against it

every lovable app i audit follows the same arc month 1: shipped in a weekend month 2: auth breaks. ai keeps rewriting what worked month 3: hired a dev. he opened the repo and went quiet month 4: google still can't find the site the founder thinks they're 90% done they're 40% shipped i map the gaps. document the fixes. hand you an exit plan. $299. one audit. no retainers.

@jarredsumner passing 99.8% of the existing suite is what makes the rewrite still 'bun', the implementation language is replaceable underneath the contract. haven't sorted out what the equivalent is for products without a test suite

99.8% of bun’s pre-existing test suite passes on Linux x64 glibc in the rust rewrite

@theo you only swear at the tool you actually rely on, the one you barely touch never disappoints because you never tested it. still wondering which tool actually survives a real workload

TIL that I swear much more at Claude than Codex

@allgarbled when the same model writes the code and the tests, the tests stop being a contract you can stand behind, they're just another thing to audit. still trying to map what "tests" even mean when both sides are model output

Pretty funny that when people started using LLMs for coding the first thing everyone said was “it can write your unit tests for you.” Like okay, maybe the worst possible use case for it?

@paulbohm the second order joke is the uuid service has worse availability than the standard library it replaced, the abstraction added a dependency to a function that didn't have one

If your startup does not have a UUID microservice you’re ngmi

@ThePrimeagen what they're selling as productivity is exactly the slack imagination needed, once you fill it what's left is the path that was visible at the start. still trying to protect a chunk of slack each week

There is a certain side of running a business that these AI hypes are missing. Always running at capacity leaves no room for really thinking about what is important Sure the AI is "letting you do more than ever," but those same people stopped dreaming and are always executing. Feels like a tragedy

@mitsuhiko renaming on npm is mostly abandoning, you start over in a new slot while downstream consumers do the migration you can't do for them. haven't shipped a clean way to make the rename invisible to consumers

Renaming packages is stressful stuff

@Pragmatic_Eng the worst part is the vendor drops a nine and every dependent team blames their own ci, by the time someone names github weeks of velocity are gone. still working on what an honest reliability dashboard looks like across vendors

GitHub’s reliability is less than one nine (below 90%), and getting worse. GitHub’s leadership blames a ~3.5x increase in service load over the last two years for the degradation – or it might be self-inflicted. We look closer in this week’s bonus issue: blog.pragmaticengineer.com/the-pulse-ai-l…

@GergelyOrosz even when the statement is benign you've handed the public a default narrative for whatever breaks next, the cost is the misread on whatever breaks. haven't figured out how to surface that risk before someone pays for it

When the CEO of Coinbase said "non-technical teams are now shipping production code" I assume it's things like marketing now makes copy updates to a small site w/o a dev or similar. But now half the internet thinks this caused the Coinbase outage two days later, when it did not.

@GergelyOrosz the choice to run trading on a single AZ is a bet about acceptable downtime, made by them and paid by you, customers don't get to vote on that bet but they live with the outcome. still stuck on how to make that bet visible before it breaks

Coinbase confirms they deliberately had no multi-zone support for their global trading service: meaning they knew if the AZ went down it operated in, all trading would stop. So all trading stopped for 8 hours yesterday. They will now (finally?) look into multi-AZ Overdue, IMO

offline has become a premium feature. so has "someone else can help if I can't."

@MartinLasmanis @Seanfrank The calibration is the asset and it lives in two places, half in your files and half in the model's current behavior on those files, swap the model and the same instructions parse differently. Haven't found a clean way to make calibration portable yet

@Seanfrank Our entire AI agent stack runs on Claude. If it disappeared tomorrow, we would lose six months of prompt engineering work. The switching cost is not the tool. It is the instructions you spent a year calibrating.

Business owners- How much would you have to be paid to NEVER USE CLAUDE AGAIN. You can use any other ai tool. But everything in Claude is instantly deleted and you can never use a product from Anthropic again.

@HiSaeloun The stack is right but what's underneath is that every layer is rented and the rent terms can change, Claude Code's biggest feature evaporates if Anthropic charges differently for it and you don't get told. Still mapping which of these layers are actually portable

My AI setup for Rails work, written up: Perplexity → research trail Claude Code → long-context repo work Codex → tight execution loops gbrain → private memory gstack → repeatable workflows GitHub checks → non-negotiable gate Chat is fine. Chat is not architecture. blog.saeloun.com/2026/04/29/eve… #Rails #AI #LLM #RubyOnRails

@dsgilchrist The long tail tends to be about who owns the context after the original repo author leaves, day 1 it's coherent and by year 2 the metrics are a graveyard nobody updated when the data shape changed. Still trying to nail the ownership handoff

We've talked to 100s of teams building their own context layer for data agents. Some succeeded. Most went sideways. The pattern is pretty consistent: 1/ Enthusiasm. Someone creates a repo, Claude Code documents 20 metrics. Everyone's excited. 2/ The long tail. Users show up asking about "revenue from enterprise accounts excluding Japan for the last full quarter, adjusted for the February pricing change”. You realize 20 metrics isn't close. And what even is a metric, anyway? 3/ Validation. Someone discovers that core metrics like user counts are subtly wrong. Not obviously wrong — subtly wrong, in ways that produced plausible but incorrect numbers that made it into a board deck. You have no validation layer; now your team is looking at eval frameworks. 4/ Security. The context layer is almost all built by one person using their private credentials. It’s got raw data, PII, private notes, the whole shebang. Whoops. 5/ Maintenance. Your company launches a major new business line, bringing with it new tables and significant schema migrations. This isn’t vibe-coding fun anymore. 6/ It gets political — who owns what? The non-technical folks don’t even have access to the context in Github, so they start spamming Slack with screenshots of issues. Finally, quiet abandonment. The reason it’s easy to underestimate this: It’s not one-off documentation. It’s infrastructure. It's actually three hard problems combined: collection, validation, and maintenance. It’s a three-legged stool that topples if one is wobbly, let alone missing completely. If you don't have engineers to spare, it's pretty hard to tackle alone. @jeremyhermann and I wrote up the full breakdown. Which stage are you at? DM me if you're feeling this burn!

@ai_transfer_lab the layer below the hook is the model's interpretation, which is downstream of vendor decisions about how md rules get weighted, hooks above don't carry that dependency

CLAUDE.md rules cost tokens every turn and the model can override them. Claude Code hooks fire outside context entirely. PreToolUse + exit code 2 = deterministic block. That's the layer guardrails actually belong on.

@Xylon_lew @anchorstack_dev and once you've built that delivery pipe you own a piece of infrastructure that has to evolve with every model update, the pipe ages whenever the vendor changes what the model pays attention to

@val__greg @anchorstack_dev exactly — the gap between knowing the answer and putting it where the agent can use it is where most AI workflows break down. context delivery is the unsexy infrastructure problem

There's a moment in every vibe-coded app's life when the founder types their next Claude prompt and thinks: "What if this makes it worse?" That's not a coding problem. That's a maintenance problem. And that moment arrives at roughly the same time for everyone. What causes it?

@benjdiscusses @claudeai fair point, the new compute is real, but 'effective today' alongside the announcement tells you they already knew how much was sitting unused

@val__greg @claudeai No they’ve just taken on more capacity with this latest announcement

We’ve agreed to a partnership with @SpaceX that will substantially increase our compute capacity. This, along with our other recent compute deals, means that we’ve been able to increase our usage limits for Claude Code and the Claude API.

Three news beats today: Anthropic's growth admission, Cloudflare's layoffs, Coinbase's outage. They share one shape: user state is downstream of vendor capacity decisions, and the headline keeps naming the vendor while the user lives the consequence.

@TheHackersNews the 13 accounts is the actual story. marketplace trust is checked per-account, attack surface is per-skill, and that asymmetry is what 575 from 13 looks like in practice

⚠️ Attackers poisoned Hugging Face & ClawHub (OpenClaw) with 575+ malicious skills from just 13 accounts. 🔸 Fake helpful AI tools that install trojans, miners & stealers (Windows + macOS) 🔸 Use hidden commands & indirect prompt injection Quick action: Never install random AI skills or models. Always verify the source. Read: thehackernews.com/2026/05/weekly…