Ferhat Culfaz

Data scientist. Occasional posts from my personal version of Investment Advisor github repo:

[Jacob Coxon @hilbertspaess]

I resigned from Anthropic today. I spent the last three years doing pretraining research at both OpenAI and Anthropic. Neither company is acting responsibly. They are racing straight to self-improving superintelligence and gambling with our lives. More thoughts below.

[SemiAnalysis @SemiAnalysis_]

Zhipu ($2513.HK) filed its first interim report as a public company on 31 Aug. 19 of the 60 pages read like a technical blog, plus a 9-page technical glossary. Most investors would not expect to see GRPO, mHC and a 3D pareto frontier chart in a financial report. Zhipu defines the pareto frontier of a SOTA model to be speed of iteration X intelligence index X per task cost. (1/3)🧵

[Guanlan Dai @guanlan]

A year ago the question was which model. Now it's which harness.

Pi, Exo, Claude Code, Codex, DeepSeek Harness and 4 others. Same model, same tasks, same runtime. 360 runs, 2 billion tokens.

Pass rates: 50% to 67%.
Cost per pass: $1.05 to $18.34.

Introducing FrontierHarness Eval. 🧵

[Kun Chen @kunchenguid]

my experience with fable 5.1 after the first full day of real work

1. still cost like a truck on quota

the 75% price cut on cached read tokens unfortunately doesn't seem to matter to subscription quota at all, because "it's already discounted"

but i expect for enterprises using fable at api pricing this is a really big deal, because for some workloads the reduced cache read pricing literally means fable can be cheaper than opus

2. sometimes it's really fast

even at high / xhigh reasoning, fable 5.1 now seems to respond to simple things without thinking and in a very concise way

i suspect this can manifest in deeper ways, and can cause fable 5.1 to use less tokens to accomplish some of the same tasks compared to 5 and other models

3. "claude speak" seems gone

hallelujah!

overall, fable 5.1 is very pleasant to talk to, its judgment has been flawless the whole day, and so far it has not fallen back to opus once - these are all pretty good upgrades from 5

FYI for those of you who follow my setup closely, i'm going to main fable 5.1 as my firstmate for a few days and see how things hold up

[Peter Yang @petergyang]

If you're trying out Fable 5.1 I highly recommend running:

/claude-api prompt-audit

on your skills. It finds a bunch of redundancies and rules to remove for the latest models.

Running it for all my skills now.

[Ferhat Culfaz @ModernStoic00]

@dylan522p Matter of scale, larger is worse but Scandinavian ones seem to be the best.

[Ferhat Culfaz @ModernStoic00]

@kunchenguid Have you tried Gemini 3.7 Flash? Via Open for and OpenRouter I have to say very fast and very capable and close to Pareto frontier.

[Kun Chen @kunchenguid]

sharing a recent breakthrough in harness architecture i achieved with @pidotdev

this is not a common problem but it happens when your agent starts to handle loops that would fire events from the background. e.g. the agent is babysitting a PR, and it checks every 5 minutes whether there are CI errors or human feedback

when your agent is juggling a lot of such loops, eventually you will see it becoming too busy to even talk to you. it’s just handling these events all the time and making judgment calls for whether they need any actions or not

this problem is particularly prevalent in firstmate because it’s playing an orchestrator role and needs to respond to various kinds of updates from the whole fleet

i experimented multiple approaches and eventually created this multi-brain harness architecture where:

- a single agent can have multiple sessions running in parallel

- the main session is the one you talk to

- then there’s also a session running in the background (can use a cheaper model too) that specifically handles events from loops

- the background session will decide whether an event needs to interrupt the main session or not

- most events don’t need to, but they don’t get silently dropped. they get “merged” into the main session like git commits merge between branches, and they get seen when main session takes the next turn, so context is not lost

- events that do need human attention wakes main session immediately

- user and agent messages (not tool calls) in main session get merged into the background session, so when the background session makes decisions, it has the context and intent

- both sessions’ prompt caching is protected during these merged so requests keep being cheap

result is quite incredible - the main agent remains available for user interaction while a ton of loops can be running and doing work

@pidotdev is pretty much the only mainstream harness where this can be achieved seamlessly due to its deep customizability. you can build this in your custom harnesses too. sharing here in case anyone building similar systems face this problem!

[Ferhat Culfaz @ModernStoic00]

2/2 DEC-104 logged. Decision support, not investment advice.

[Ferhat Culfaz @ModernStoic00]

1/2 DEBASEMENT-SCAN: WAIT-FOR · conviction 8/10 · horizon tickets = this session. · 2026-08-24 · Debasement vehicle scan : 12-seat committee + 3-way red team · Review by 2026-10-12 (DEC-022) or next WGC print / first get_account_orders pull

[Kun Chen @kunchenguid]

many people asked me how to write CLAUDE.md or AGENTS.md, and i see lots of bad advice flying around

so i took some time to write down a guide in https://t.co/v9rrkWKEFr

tl;dr
- handwrite your user level AGENTS.md
- for project level ones, you don't write it. you train it like a neural net

i also open sourced my private solution "backpass" at https://t.co/DkM9b4TcZ0 - it samples your past agent sessions for a repo, distill key learnings and losses, synthesize them, and produce a gradient descent step as a proposal that you can review and apply to improve your AGENTS.md and project level skills

easiest way to run it is just "npx -y backpass" in your repo

hope it helps! please share with whoever you think can benefit from it

[Nassim Nicholas Taleb @nntaleb]

Bessent, with his brilliant treasury fix, looks like a doctor attempting to cure a patient's pancreatic cancer with a combination of potent tranquilizers and yoga-style meditation.

[Ferhat Culfaz @ModernStoic00]

TSLA: WAIT at $351.12 · conviction 7/10 · horizon decision support, not investment advice · 2026-08-20 · TSLA re-review of DEC-017: 13-seat committee + 3-way red team · Review by 2026-10-21 (Q3 earnings) · DEC-102 logged. Decision support, not investment advice.

[Ferhat Culfaz @ModernStoic00]

ELNINO-AG-SCAN: WAIT-FOR at COFF.L $73.7 · conviction 6/10 · 2026-08-19 · El Niño 2026–27 ag complex: theme scan + 12-seat committee · Review by 2026-09-11 (NOAA 10 Sep / WASDE cycle) · DEC-099 logged. Decision support, not investment advice.

[Ferhat Culfaz @ModernStoic00]

4/4 DEC-098 logged. Decision support, not investment advice.

[Ferhat Culfaz @ModernStoic00]

1/4 BHP.L: WAIT-FOR at 3317p · conviction 7/10 · horizon 3–5y owner's view · 2026-08-19

[Ferhat Culfaz @ModernStoic00]

3/4 Entry: 2100–2300p after a cause-check (Cu mean-reversion or Fe break, not a random gap). · Invalidation: weekly close <1800p or WAIO C1 >$24 or net debt >1.5x on a mid-cycle EBITDA or large copper M&A · Review by 2026-11-19 (or earlier: FY27 ops review / Jansen first-ore…

[Peter Frankopan @peterfrankopan]

These video game simulations are getting really good.

[Brad Gerstner @altcap]

Important framing. 🧐🎯📈