Stefan "Chefo" Dimitrov

In IT since 1997. I integrate AI into real businesses at Data Crafted Consulting. Blunt takes on AI models, dev tools, and what actually ships.

[Stefan "Chefo" Dimitrov @elchefo]

@glazecl Fair point. A token only tells you which app is asking, not what it should be allowed to read, so scoping per client is the other half.

[Stefan "Chefo" Dimitrov @elchefo]

What I'd flag in Yogurt, the open-source Granola alternative on HN: its MCP access has no per-app token. Turn it on and any app running under your user can read your meeting transcripts, not just the assistant you added. Off by default; audio and AI stay local.
#MCP #Privacy

[Stefan "Chefo" Dimitrov @elchefo]

Stela (first release Nov 11) aims to run macOS 27 on x86 PCs as a fake M2 Mac mini. I'd watch the GPU path more than the CPU translator: Metal shaders compiled to Mesa's NIR for NVIDIA. Without Metal, macOS is barely usable. README: slow, may not boot yet.
#macOS #Mesa

[Stefan "Chefo" Dimitrov @elchefo]

Terence Tao is rebuilding QED, his 2018 logic textbook-game, with a coding agent (working title "Lean Grimoire"). He says the old UI was capped by his graphics skills. The math was never the bottleneck; the frontend was, and that's what the agent took over.
#LeanProver #AIAgents

[Stefan "Chefo" Dimitrov @elchefo]

@Fanfulladev Agreed. Until there's a per-app token I'd keep it off on any machine where you don't trust everything running under your user.

[Stefan "Chefo" Dimitrov @elchefo]

Huawei open-sourced MACaQ (Apache 2.0, Nov 10). The useful part is MACaQ-R: plain MACaQ GGUFs need Huawei's runtime, -R runs on stock llama.cpp. Its Qwen3.6-35B-A3B Q4_K_XL: same 22.36 GB as Unsloth's, ~15% lower KL to FP16. Huawei's numbers.
#LocalLLM #llamacpp

[Stefan "Chefo" Dimitrov @elchefo]

Apache httpd 2.4.71 (Nov 11) fixes critical mod_http2 RCE CVE-2026-119618 in 2.4.55-2.4.70. Apache credits Anthropic's OSS Scanner (reported Oct 26) and Google researchers (Nov 3): two reports of one bug before a fix, which shipped 16 days after the first.
#Apache #InfoSec

[Stefan "Chefo" Dimitrov @elchefo]

Google's Gemini Decisions API scores closed-set answers in one forward pass: up to 1,000 questions per input, $0.10/1M input tokens, output free. Its latency claim (150-250 ms, nearly flat from 1 to 100 questions) is for contexts up to 16K tokens, not the 256K max. #Gemini #LLM

[Stefan "Chefo" Dimitrov @elchefo]

Mozilla simulated what Fingerprint Pro, Signifyd and Akamai scripts measure, using data from nearly 25M Firefox users. Firefox can now report CPU count as just 4 or 8 and hides taskbar/dock size. Mozilla says precision dropped substantially; trackers can adapt. #Privacy #Firefox

[Stefan "Chefo" Dimitrov @elchefo]

Apple Support Community is read-only and closes Nov 20. The Verge: Apple hasn't said what happens to the posts (the forum dates to 1998) and presumably they'll disappear. Any https://t.co/xwgEm2pZXd link in a ticket or runbook is at risk. #Apple #SysAdmin #MacAdmins

[Stefan "Chefo" Dimitrov @elchefo]

Pietro Miotti says 48 general agents, no special model or classifier, went through ~1,200 km² of NASA lidar images, following a precise guide. INAH's records show no registered site at the flagged hilltop. He calls it "not official yet"; no field check so far.

#AIAgents #Lidar

[Stefan "Chefo" Dimitrov @elchefo]

Kate's blog (Nov 10): voltage-glitching the GA100 fuse controller on NVIDIA's CMP 170HX mining card lets the author burn extra fuses. The 8GB/70-SM card gets 64-80GB and ~120-126 SMs. OpenCL FP16 FMA: 85.9 TFLOP/s vs 72.5 on a retail A100.

#NVIDIA #HardwareSecurity

[Stefan "Chefo" Dimitrov @elchefo]

FAISS 1.16.0 (Nov 10) ships Meta FAIR's e-means. Authors say it matches Lloyd k-means' 25-epoch loss in 1.4-2.6 epochs, 10-45x faster on GPU. Their caveat: the tuning laws target high-dimensional data, not tiny or very low-dimensional sets. #FAISS #VectorSearch

[Stefan "Chefo" Dimitrov @elchefo]

Railway's breach report (Nov 10): env vars used envelope encryption, root key in a cloud KMS that can't leave it. The attacker didn't need the key. A file on an orchestration machine held a long-lived credential they used to have KMS decrypt variables. #DevSecOps #CloudSecurity

[Stefan "Chefo" Dimitrov @elchefo]

Baidu says ERNIE 5.5 Flash Preview weights will be open-sourced under Apache 2.0 "soon." Its Nov 5 Base license excludes the EU, UK and South Korea. Apache 2.0 has no territory carve-out, so EU teams get what Base denied them, if it ships that way.

#OpenSource #ERNIE

[Stefan "Chefo" Dimitrov @elchefo]

Anthropic's roadshow says Mythos got Claude's prompt processing running on AMD MI355X in ~3 days, not the usual 6-12 months. That's one stage of serving. Analyst Ben Pouladian: part of a model on AMD isn't stable, cost-effective deployment. 3 days is bring-up.

#AMD #AIInfra

[Stefan "Chefo" Dimitrov @elchefo]

Apple's unannounced Harbor app pools unified memory across Thunderbolt 5 Macs for local models: "no manual configuration, no command line" (9to5Mac leak report). It's in private preview under Claris, Apple's FileMaker unit, which suggests business users first.

#Apple #LocalLLM

[Stefan "Chefo" Dimitrov @elchefo]

Gemini 4 Carbon costs $2/$10 per 1M tokens until Nov 30, then $4/$20. Google's launch post also claims #1 on three benchmarks at "more than 2X lower cost per task" than the next best external model, but never says which of the two prices that figure uses.

#Gemini #GoogleAI #LLM

[Stefan "Chefo" Dimitrov @elchefo]

Epic's UE 5.9 MegaJam Preview is "the first Unreal Engine release built primarily with agentic workflows". It had "about a week of development and a shorter QA pass than usual", timed for the Nov 12-15 AI MegaJam. Parts of the web renderer are prototypes.
#UnrealEngine #AIAgents

[Stefan "Chefo" Dimitrov @elchefo]

Expo's Charlie Cheever built Exact, pre-1.0: UI in new language Contract, "no webview and no JavaScript bridge in the UI". Pitch: "A whole app fits in a context window."
README's agent test: a todo app, 57 min, 25 on first builds, warm caches.
#TypeScript #RustLang #AIagents