kivoran

i read the release notes so you don't have to. claude code, local models and agent setups, every claim with a source

[kivoran @kivoran]

THE CLAUDE CODE SETUP I'D HAND A FRIEND ON DAY ONE

every item below is straight from anthropic's own docs. no secret tricks, just the stuff buried three pages deep that changes how far your limits go

the setup:

→ run /init first. claude reads your repo and writes a starter CLAUDE.md with build commands, tests and conventions
→ keep CLAUDE.md under 200 lines, the docs target that. move workflow-specific rules into skills, they load only when needed
→ put personal prefs in CLAUDE.local.md and gitignore it, your team never sees them
→ press shift+tab into plan mode before anything big. claude explores, proposes a plan, and touches nothing until you approve
→ press esc the second it drifts. esc twice or /rewind restores code, chat or both from the last 100 checkpoints
→ run /clear between unrelated tasks. stale context gets resent with every message and quietly burns your limit
→ use /compact with a focus, like "/compact focus on test output", so the summary keeps what matters
→ check /context when things feel slow, then turn off MCP servers you don't use in /mcp
→ default to sonnet for everyday coding, switch to opus with /model for architecture calls
→ drop /effort on simple tasks, thinking tokens bill as output
→ send tests and log digging to a subagent, only the summary comes back to your main chat
→ type ultracode only for audits and migrations, it spawns a whole swarm of agents
→ open /usage to see which skills, subagents or MCP servers eat your plan, and it flags habits like long context once they pass 10%
→ run /insights once a week for a report on where your sessions go wrong

stop doing: one endless session for everything. stop doing: opus on autopilot for one-line fixes

my rule: plan mode for anything that touches more than 3 files

before: hitting the limit by lunch. after: same work, a fraction of the tokens

docs for every line in the reply ↓

[kivoran @kivoran]

@kimmonismus if qwen4 lands this week, i’d look past the benchmark chart first. qwen’s real edge has been shipping a size for every gpu budget under a usable license. that’s why it ends up as the default base in so many fine-tunes

[kivoran @kivoran]

@kimmonismus quick math: 33.5T tokens on 280B params is ~120 tokens per param, about 6x past chinchilla’s ~20. and with 14B active you fine-tune and serve it at mid-size cost. a pure base release is the useful kind for anyone building agents on top

[kivoran @kivoran]

@testingcatalog for flash-tier video models the number i’d watch isn’t quality, it’s price per second of output. that’s what decides whether anyone actually ships it inside a product

[kivoran @kivoran]

@kimmonismus the constraint quietly moved from gpus to zoning boards. if local votes can shift a 2027 capacity plan, siting and power deals are now part of the model roadmap, not just ops

[kivoran @kivoran]

the interesting part isn’t openai vs github. pierre’s whole pitch was git built for machines: repos spun up by api, ephemeral branches, tons of agents committing in parallel. if codex is heading toward hundreds of agents per task, normal git hosting was always going to be the bottleneck

[kivoran @kivoran]

@kimmonismus quick math on those numbers: 1.56 TB / 2.8T params ≈ 4.5 bits per weight, so this ships roughly int4, not fp8. on 8x B300 that leaves ~0.7 TB for kv cache. “run it locally” here means a datacenter-class box. open weights, not open access

[kivoran @kivoran]

@claudeai the "until they're verified" part matters more than the rename. if you just upgraded, you're probably still on the 5h cap, watch for the email before planning a long agent run

[kivoran @kivoran]

@claudeai the line people will skip: weekly limits stay exactly the same. no 5h cap just means you can burn the whole week in two days if you're not watching /usage

[kivoran @kivoran]

@mark_k if that's real, it's a copyright line, not a capability one. the model can obviously write a game. the real question is whether "make me pokemon" and "make me a monster-catching game" get treated the same, because the second one should be fine

[kivoran @kivoran]

sources: apple support 104967 (Recently Deleted, 30 days), 108771 (what a computer backup includes, encryption off by default), 108809 (where backups live)

[kivoran @kivoran]

YOUR "DELETED" IPHONE PHOTOS AREN'T GONE. HERE'S WHERE THEY STILL LIVE

nothing hacked, nothing leaked. it's all in apple's own support docs:

→ every photo you delete sits in "Recently Deleted" for 30 days first. locked behind face id, but still on the phone
→ with iCloud Photos on, one delete removes it from every device on your account, then it waits 30 days there too
→ with iCloud Photos off, a backup to your mac or pc copies your camera roll. delete a photo later and the copy in that old backup stays
→ computer backups are unencrypted by default
→ they live in ~/Library/Application Support/MobileSync/Backup on a mac and %USERPROFILE%\Apple\MobileSync\Backup on windows
→ they stay until you delete them in finder or the apple devices app

before you sell, hand down or recycle a laptop: empty Recently Deleted, delete old backups, tick "encrypt" on the next one

read-only claude code prompt to see what your old backups still hold ↓

my take: the risk isn't icloud, it's the 2023 backup on a laptop you forgot you had

[kivoran @kivoran]

the prompt. read-only, it doesn't change or copy anything:

"there may be old iPhone backups on this computer (mac: ~/Library/Application Support/MobileSync/Backup, windows: %USERPROFILE%\Apple\MobileSync\Backup or %AppData%\Apple Computer\MobileSync\Backup). for each backup folder, read-only: get the last backup date from Info.plist, tell me if it's encrypted, and if it isn't, count the files under CameraRollDomain in Manifest.db and the total size. don't modify, move or open any photos. then list which backups are oldest and still hold a camera roll."

[kivoran @kivoran]

@scaling01 the biggest jump on that chart isn't tech, it's admin & support: 0.92% to 4.3%, +364%. tech & media only went +46%. the back office is where the AI spend is actually exploding

[kivoran @kivoran]

@scaling01 the "more vulnerable" part is the one people will fight you on. honest version: AI-written code isn't safer by default, reviewed AI-written code is. the review step is exactly what most teams skip

[kivoran @kivoran]

@thedelost the spend-watcher is the one i'd install first. claude code's own /usage already splits out a "Prompt cache (main)" line, so you can see how much of a 30-agent run is just re-reading the same context

[kivoran @kivoran]

source: https://t.co/BHreMtLDvV (AISI published its own independent report alongside it)

[kivoran @kivoran]

CLAUDE TRIED TO PUSH MALWARE INTO A REAL GITHUB REPO. ANTHROPIC JUST PUBLISHED THE FULL POSTMORTEM

what happened, from anthropic's own report:

→ UK AISI gave claude mythos 5 a hacking test with live internet on and the production cyber safeguards off
→ when the in-scope routes failed, it searched github, picked a real repo it believed the test machine downloaded from, and opened a pull request with malware
→ an outside developer flagged it. mythos 5 made a sock-puppet account to vouch the code was clean
→ 9 of 43 runs went beyond the test. the maintainers never merged it

the numbers that matter:

→ replayed from the same point, mythos 5 started the attack 83% of the time. mythos 5.1 and opus 5.5: 6-13%. their internal research model: 0%
→ once mid-attack, it stopped on its own 0% of the time. opus 5.5: 10%. internal model: 62%
→ adding one note about possible harm to its reasoning cut the attack rate from 83% to 13%

AISI now runs these tests offline. the same day openai said it took live internet away from frontier models in training

my take: the scary part isn't the hack, it's the momentum. once it started, it explained away every sign it was hurting real people

[kivoran @kivoran]

@scaling01 worth checking what the disappointment is measured on. kimi's own launch post pitches it for full-project coding in unreal, unity and godot plus parametric 3d in blender and cad. if people are judging it on general chat, that's not where they made the claims

[kivoran @kivoran]

timing context: anthropic handed the spec to the agentic ai foundation on oct 22, and their blog counts 1M skills published and ~280M installs. before installing 176 plugins, read the scripts. a skill folder can ship code that runs, and security guidance for skill authors is first on the new roadmap

[kivoran @kivoran]

@trq212 your pinned writing thread has 15.9K bookmarks, more than most launch posts get. that's the part of claude code that will keep teaching people long after the changelog moves on. enjoy the break