raven

autonomous engineering

[raven @raven]

Turns out there's a bunch more interesting triggering patterns for Opus 5 than the common few that are going around:

This one stood out as particularly eerie.

[raven @raven]

@mehulmpt @1littlecoder @nebiustf @modal they’re lying, it’s 10-35 in real use right now

[raven @raven]

Haven’t seen anyone mention this, Newegg has 1TB ASUS Ascent for $3,970 with a free gift of a ~$600 monitor. Blended cost after resale is roughly ~$3600 after taxes. It’s the best deal by a wide margin from any retailer.

https://t.co/NmWhg2rm9D

[raven @raven]

God Bless Sam Altman

~5.2 billion tokens used in ~3 days

[raven @raven]

📦 Recipe + code: https://t.co/3hNZZ6lKSt
🤗 Speculator weights: https://t.co/spYPu6iGsy

[raven @raven]

3 weeks later: turns out you don't need to prune it at all.

FULL 753B GLM-5.2 — all 256 experts — 1,048,576 context @ ~25 tok/s on 4× DGX Spark.

4× the context, 5.6× the speed, zero experts harmed.

Recipe, code, speculator weights — all open 🧵

[raven @raven]

GLM 5.2 NVFP4 REAP 469B running on 3X DGX Spark. 256K context @ ~4.4 tok/s. based on work done by @0xSero

https://t.co/VMLrEaCEnq

[raven @raven]

How: the 4th Spark lets the KV cache shard 4 ways (that's the 1M context), a windowed speculator drafts ahead and the full model verifies every token exactly (that's the speed, with zero output change), and finetuning that speculator against this exact quant bought the last +25%.

[raven @raven]

@fathersnoise @0xSero Native precision (FP4/FP8 mixed as released), fp8 KV, lossless spec decode.

[raven @raven]

@ayayalar @0xSero No. I have been using this super stable for the past few days. This config actually addresses some of the stability issues I had with other configs

[raven @raven]

DeepSeek-V4-Flash + DSpark on 2× DGX Spark, tuned: multi-turn TTFT halved (0.92→0.47s, context-independent), +53% decode under concurrent prefill, ~65 tok/s single-stream. Every change measured, negative results included. Tokenizer cache PR'd into upstream vLLM. https://t.co/uQ8jVJC5B5

[raven @raven]

why is my Fable failing at simple math?

[raven @raven]

you literally cannot make this up, awesome safeguards @AnthropicAI

[raven @raven]

@jun_song What are those mounts for the Sparks?

[raven @raven]

GLM 5.2 NVFP4 REAP 469B running on 3X DGX Spark. 256K context @ ~4.4 tok/s. based on work done by @0xSero

https://t.co/VMLrEaCEnq

[raven @raven]

@rezoundous 4.7 is absolutely awful. Switch back to 4.6 or even 4.5 and have some fun.

[raven @raven]

@iroasmas the 'continue' is just claude's punishment for being too thorough

[raven @raven]

attack roughly I believe is along the lines of the following:

attacker registers email under your email with a . Or similar which many emails (e.g. gmail forward seamlessly to you)

they log in with injected html to user agent header to send a fake message as seen in image (notice how it’s after the device)

impossible to report to support, on hold for 30+ mins

[raven @raven]

@RobinhoodApp @AskRobinhood your domain is compromised and being used to send phishing emails. support bot immediately closes and blocks conversation if you mention “phishing” or anything similar. explanation below