Karl

@karlkfi

SF Tech Gamer Car Nerd. AI Infra Tech Lead @ AMD (opinions my own). Previously Google, Cruise, Mesosphere, Pivotal.

145 Siguiendo539 Seguidores744 Publicaciones

YouTube YouTube Bluesky

gamer

¿Es tu perfil?Reclámalo y karlkfi.dirty.social es tuyo.Reclamar perfil

Karl@karlkfi· 7d

Welp… guess I’m not gonna get that 700 day Reddit streak cheevo. Counter reset because I worked all day then passed out from not sleeping the previous night. Maybe this is a sign.

Karl@karlkfi· 8d

I posted the cache chart in a previous thread in the quote chain, but stopped posting it because it's just silly. The tokens reported above are non-cache. Here's the latest one with reads & writes.

Area chart titled "Cumulative cache traffic" covering May 17 to Aug 3, 2026. A blue area for cache reads (replayed context) climbs steadily to 17.92 billion cumulative tokens, while an orange sliver along the bottom for cache writes (writes once) reaches just 345 million — reads outnumber writes by roughly 50 to 1. Dashed vertical lines mark plan upgrades: "Pro → Max 5x" around May 24 and "Max 5x → 20x" around Jul 6, with a green line marking "mac-2 begins" around Jul 26, after which the reads curve steepens sharply. A footnote says the shaded band (May 16–18) is estimated from archived sessions.

Karl@karlkfi· 9d

Surprisingly, Opus 5 runs 21% cheaper per message than 4.8. But a beefy new toy and a growing addiction meant a lot more tokens spent. 96 hours at the keyboard became 286 hours of session time in 9 days. ~3 at once, peak 13. Parallel dispatch is the trick. One task, one session, one branch.

Two stacked charts, July 26 to August 3. Peak concurrent Claude Code sessions per day, 8 rising to 13, with mean concurrency far lower at 2.0 to 3.9. Below, 96 hours at the keyboard against 286 hours of session time summed across concurrent sessions.

Karl@karlkfi· 9d

Day 80 of building a Kubernetes operator with Claude Code: 1,654 commits 1,876 tests 424M tokens 198k lines authored ~2,140 tokens/line, flat since day 70 Daily token spend jumped 2.9× after I got my new M5 Max 128GB MacBook Pro!

Three stacked charts, May 19 to August 3. Cumulative tokens (424M) far above lines authored (198k) on a log scale; the lines broken down by type, tests and docs largest; and cost per line rising from 406 to 2,140, mostly before mid-June.

Karl@karlkfi· 18d

Only a thin band of gold wrapping the horizon over here from the train. 🚂 🌲☀️

Over aggressive autofocus focusing all 4k pixels on the water spots of the double pane train window while an orange haze melts into a purple sky above some blurry coniferous trees.

Karl@karlkfi· 19d

The headline number undersells it. Counting cache reads: 10.5B tokens. 227M written to cache, each replayed ~45 times. The cache is the only reason this is affordable.

Daily token usage by model, May 16–Jul 24, stacked bars in millions/day. Two dashed lines mark plan upgrades: "Pro → Max 5x" on May 23, where gold Sonnet 4.6 hands off to pink Opus 4.7 then blue Opus 4.8 (peak ~13M on Jun 15), and "Max 5x → 20x" on Jul 5, after which usage holds ~4–7M/day. Green Fable 5 appears June 9 and takes a growing share through July, with some late-July days mostly green. First three bars hatched: estimated from archived sessions.

Karl@karlkfi· 19d

Day 70 of building a Kubernetes operator with Claude Code: 1,221 commits 1,394 tests 284M tokens 137k lines authored ~2,050 tokens/line — barely rising now Scaffolding was cheap. Logic, tests, and review are where the tokens went. Now the conventions are paying rent.

Three charts, May 19–Jul 24: cumulative tokens (284M) vs lines authored (137k) on a log scale, gap ≈2,068 tokens/line; a stacked breakdown of the 137k lines (Go code 28k, tests 44k, docs 42k, YAML 9k, scripts 15k); and tokens-per-line over time, climbing steeply from 406 to ~1,900 in seven weeks, then flattening to 2,068.

Karl@karlkfi· Jul 6

My Z would like a chat.

Karl@karlkfi· Jul 3

Day 48 of building a Kubernetes operator with Claude Code: 965 commits 976 tests 188M tokens 98k lines authored ~1,900 tokens/line and rising The cost-per-line climbs ~5× as the work shifts from scaffolding to logic, tests, and debugging.

Karl@karlkfi· Jun 21

I had way too much fun marking this! "AI generated", but frame-by-frame with code.

Karl@karlkfi· Jun 20

35 days building github-actions-gateway with Claude Code: 📈 115M tokens spent 📦 60K lines authored 🔁 786 commits 🐹 28k lines of Go 📚 20k lines of Documentation 🧪 547 test cases Each new line cost more tokens! Most of the work is reading, writing, testing & reviewing, not cranking out code.

Karl@karlkfi· Jun 8

Claude's cache doing heavy lifting here. 1.89B cache reads. 42M cache writes. 22 days. Every token I cached got replayed ~45 times!

Stacked cumulative area chart, May 19–June 8. A large dark area (cache reads) climbs to 1.89 billion tokens. Cache writes form a thin band along the baseline totaling 42 million, marked with an arrow. Cache reads vastly outnumber writes.

Karl@karlkfi· Jun 8

Upgrading to Max made using Opus for everything viable. Every gold bar before the line is Sonnet 4.6 on Pro. The second Max kicked in, I switched to Opus — and rode the 4.7 → 4.8 upgrade straight through. Daily token use jumped with it.

Stacked daily token-usage bars by model, May 19–June 8, with a "Pro→Max" line on May 23. Before it: all Sonnet 4.6 (gold). After: Opus 4.7 (purple) then Opus 4.8 (blue). The model flips from Sonnet to Opus right at the upgrade.

Karl@karlkfi· Jun 8

22 days into this side project with Claude Code... Day 7 → Day 22: 10M → ~56M tokens 232 → 617 commits 15.5k → 20.9k lines of Go 2.3k → 4.2k lines of comments 14.3k → 18.3k lines of Markdown 1.5k → 1.8k lines of YAML (without CRDs) 269 → 393 tests Mostly Sonnet → Mostly Opus Pro → Max

Line chart of cumulative growth since the day-7 tweet (1× baseline). From May 19 to June 8, tokens used (gold) climbs to 5.3× while git commits reach 2.7× and tests and lines of Go stay near 1.4–1.5×. Token usage grew much faster than the codebase.

Karl@karlkfi· May 25

Helios racks are coming. - CPUs: 18x EPYC “Venice” (2nm, 256 physical cores each) - GPUs: 72x MI455X (432 GB HBM4 each) - NICs: 36x Pensando Vulcano (800 Gbps each) - RAM: 288x DDR5 (1.6 TB/s per-socket) Venice production is ramping up now. www.tomshardware.com/tech-industr...

AMD CEO Lisa Su holding a Venice CPU in front of a Helios tray

Karl@karlkfi· Apr 1

Muse Tix Acquired

Karl@karlkfi· Mar 5

Wow. This lineup is wild. #AftershockFestival aftershockfestival.com

Karl@karlkfi· Oct 6

Joining AMD in May is looking pretty clairvoyant in retrospect.

AMD stock valuation for the last 2 years

Karl@karlkfi· Jun 19

New hood, who dis?

Fujimura Auto Rocket Dancer Vented Bonnet on 370z

Karl@karlkfi· Jun 15

At brunch today

Karl@karlkfi· Jun 2

The Keys in town

Karl@karlkfi· May 13

I've joined AMD! The future for me looks a lot like the past: building platforms for internal developers. Kubernetes, containers, virtual machines, bare metal, hybrid, multi-cloud... anything AMD needs to build, test, and ship GPU drivers, AI tools, inference, training, demo clusters, you name it!

Karl@karlkfi· May 1

First time ever.

Karl@karlkfi· Apr 2

Frieren is great. What if the elf that went on the hero’s quest and helped defeat the Demon King outlived the rest of her party? What would her next adventure be? On Crunchyroll and Netflix.