Quant Cat (@0xQuantCat) header
Quant Cat (@0xQuantCat) profile picture

Researching AI Context, Efficiency & Architectures | LLMs, Agents, Protocols My profile is a journal

Always nearby
Joined March 2026
The transformer is now the baseline. What actually separates good models from great ones is almost always better data, smarter scaling choices, effective post-training, and serious systems engineering. Feels too simple yet?
1
9
If you want the single clearest explanation of how modern LLMs are actually built, watch this Stanford lecture: teapawt.net/raytar/status/20684625…
My friend applied to 200 tech jobs in two years. No CS degree. No callbacks. Last month Anthropic offered him $750,000. All because of one Stanford lecture. Free on YouTube. One hour. A professor explains how ChatGPT actually works. Not the Twitter version. The real one. He watched it in bed. Paused it eleven times. After that hour he told me something I didn't believe. "It's embarrassingly simple." Three days later he applied to Anthropic. Every single question they asked him, he knew from that video.
Readers added context they thought people might want to know
Anthropic's engineering interviews require coding challenges, system design, past experience discussions, and alignment checks; $750,000 offers target senior/lead roles with substantial expertise, not entry-level hires. interviewing.io/anthropic-inte… anthropic.com/careers levels.fyi/companies/anth…
16
Absolutely important detail. An agent won’t know when it has more than enough context. Extra context doesn’t just raise token cost; it can trip up the task if activation is loose. Runtime context governance and source-verified admission fixes this. Load what’s needed, prove why, expand only when evidence says so.
Because codex is so good at adhering to your skill files, you have to be very intentional about how you word the description, or they can trigger more often than necessary. The coolest thing is having codex run evals for skill activation using sub agents!
21
I want a “subagent of the day/week/month” wall with their picture in a frame
today is unhinged codex feature request day. you know what to do
27
Agents need memory that retains facts, retrieves selectively, updates context over time, and stays human-inspectable (our biggest bottleneck). LLMs know file systems deeply from trillions of tokens on Unix commands like ls, cat, grep, mkdir. So filesystems provide durable, revisable context with names, hierarchy, timestamps, and diffs that agents reason about naturally, but here we skip over runtime context governance. For working sets like plans, notes, and logs, filesystem interfaces offer shared legibility for agents and humans over custom schemas, but lack something important.
24
Merging OSS AI models is an absolute hidden beast (see Fusion by @OpenRouter) Once such fused model that runs recursively in self-improvement loops it’s game over. Closed models can only improve when their company updates them. This decentralized, weight-level iteration is what enables the rapid, compounding progress people like @TheAhmadOsman predict.
Replying to @TheAhmadOsman
How does open source achieve continuous learning
32
A seamless handoff of coding threads between local laptops and remote hosts, with Codex orchestrating the entire process? You can begin work locally for frontend QA and dev ergonomics, transfer threads including git state and uncommitted changes to a remote box before closing their laptop, and resume later without manual setup. #Codex is listing active threads, creates worktrees, transfers files, and confirms successful activation on the remote side while preserving ongoing sessions. Damn. Your turn, @claudeai
Codex can now hand off threads between local and remote hosts. Start work on your laptop, send it to a remote box before you close the lid, bring it back later. And yes, Codex can orchestrate the handoff for you.
39
The Second Renaissance
What's coming after artificial intelligence?
1
44
Multi-loops are around the corner. Wait for one bigger TL to drop in casually, that by the way they don’t loop anymore, their loops are looping other loops iteratively!
I’m not really into loop yet because the way I work is not really compatible with it?? (?) But I’m a lot into one coordinator/orchestrator thread per project which have all the context. Right now I asked to spawn worktrees and threads for all my draft open PR review them.
37
Solve your OWN problems. If your made your life easier, and you start actually using it, release the solution. Chances are other people have the same problem as you, and they will always look for existing solutions first. Let them find yours!
You're starting from zero. No audience. No money. No network. What's the first thing you do?
51
I was recently hit by a shadowban, my account was new and I was a bit too eager... If you happen to find this article because you suffer from the same problem, give it a read. If you aren't, still give it a read, the best action is prevention! Read how to avoid the shadowban:
21
Fable 5 was two Opus 4.8 in a tench coat (called it)
Introducing the Fusion API, the smartest compound model in the market. Fusion achieves Fable-level intelligence at half the price. How it works 👇
47
It’s crazy that Fable 5 and Mythos 5 is still cheaper than what Opus 4.1 was
26
.@claudeai Fable 5 used ~1.77x more tokens and took ~1.81x longer than Opus 4.8, leading to ~3.6x higher cost, but produced noticeably better results. I adjusted Fable 5 to match Opus 4.8 on the same basis assuming linear scaling for a fair apples-to-apples comparison on token spending or time spending. Cost scales with tokens naturally, and time is normalized proportionally as a proxy for generation effort. When adjusted on token spending basis of 38.9k tokens: F5 (Norm): $1.90 | 38.9k T | 8m 22s O4.8 (OG): $0.93 | 38.9k T | 8m 10s Fable 5 would cost ~2.04x more and take almost the same time. Huh. When adjusted on time spending basis of 490 seconds: F5 (Norm): $1.85 | ~38.0k T | 8m 10s O4.8 (OG): $0.93 | 38.9k T | 8m 10s Fable 5 would cost ~1.99x more and produce nearly the same token count. Huhhh… So when put on the same token or same time footing as Opus 4.8, Fable 5 is still roughly 2x more expensive. The original 3.6× cost gap was mostly driven by Fable simply generating more output, a longer/more detailed code for the sims. This normalized view highlights the pure price-per-token (or price-per-second) difference while keeping the quality edge noted in the test. If you ask me, Fable 5 is just Opus 4.8 on 2xhigh. Not a good look
New Fable 5 beats Opus 4.8 on real world physics simulations We gave both models the same three prompts and asked them to build self contained HTML5 sims with real physics and no libraries: 1. Chaotic double pendulum 2. Galton board 3. Water in a spinning drum (WCSPH) Generation cost Fable 5: $3.35 on 68.7k tokens, time 14m 47s Opus 4.8: $0.93 on 38.9k tokens, time 8m 10s Fable clearly did better on the water simulation, producing a much more solid and continuous body of water. Opus left larger gaps near the walls, scattered particles around the scene, and struggled to keep the fluid stable.
1
1
1,017
Fable 5 needs to be compared to Opus 4.8 on a same token or time spend basis Otherwise the comparison is not fair, having the cost/time advantage producing better results
67