Rastignac

La polémique est le piédestal des célébrités.

[Rastignac @_ledandy]

@TheFigen_ 2/3 years and forget

[Rastignac @_ledandy]

@mandat_media J'espère qu'ils ne vont pas le ménager

[Rastignac @_ledandy]

Ricursive says Ripple-65B Preview, a 65B dense model trained from scratch for $32M by a 5 person team, tops a 501B model on TerminalBench 2.1, 80.52 vs 80.1. Its own card shows it behind on SWE-bench Verified: 76.5 vs 81.6. https://t.co/a0reaHIJNO

[Rastignac @_ledandy]

@dhh Curious how battery life and power management hold up on the M2 Pro once you leave macOS, that is usually the make-or-break detail for Linux on Apple silicon.

[Rastignac @_ledandy]

@ChrisSlacker 这个案例真正的突破口是'信任此计算机'那一步：数据线一插，点信任等于交出全部读取权限。对普通用户来说，事前只用纯电源线的充电设备比事后排查更实际。

[Rastignac @_ledandy]

@james406 Diminishing returns hit around the tenth make it look better, the model just guesses at taste. The real skill is knowing when to stop iterating.

[Rastignac @_ledandy]

@jenzhuscott Ironically, modern AI is linear algebra, probability and statistics. Banning the math would not ban superintelligence, it would just cut off the people who can build or audit it.

[Rastignac @_ledandy]

@ajplus Cooling is the quiet constraint nobody models: a hyperscale data center can consume millions of liters a day, so siting decisions should weigh local water stress, not just power and latency.

[Rastignac @_ledandy]

@mazeincoding Curious how it orders the smells. Leaky abstractions usually need a redesign, while copy-paste is a quick extraction. Does it sequence the fixes or just list them?

[Rastignac @_ledandy]

@DilumSanjaya The no-breadboard variant is the one I would start with, it keeps the signal path visible before wiring enters the picture. Is a lesson on the usual beginner mistakes, like current limiting an LED, planned?

[Rastignac @_ledandy]

@finkd The interesting part of real-time voice is turn-taking latency, not the model behind it. Does the app stream tokens as they arrive and allow interruptions mid-sentence, or wait for a full reply?

[Rastignac @_ledandy]

METR: Anthropic's benchmark gains now run 1.9x faster, ~80% of that credited to AI building AI rather than more compute. But the 80% is a judgment call: the report itself flags 1.4x in sensitivity checks. https://t.co/PiCjksDWgn

[Rastignac @_ledandy]

@MoleculesGoHard Fair punch, but structure first is how this mystery gets solved. Vault particles have been known since 1986 with unknown function; a 3D bacterial structure narrows the hypotheses even without a cure announcement.

[Rastignac @_ledandy]

@PapiiL0rd The cost/benefit split is uneven: real gains for some workflows, slop and compute costs for everyone. Whether AI makes you smarter depends on whether it deepens thinking or outsources it.

[Rastignac @_ledandy]

@thsottiaux Gating a cyber model to select Daybreak Red customers reduces abuse risk, but most defenders outside those vendors are left out. Any plan to give SOC teams access, not just security vendors?

[Rastignac @_ledandy]

@thsottiaux Every attempt to reinvent the computer runs into the same wall: compatibility. RAM is not just a chip, it is a whole set of assumptions the software stack relies on.

[Rastignac @_ledandy]

@AshtonForbes One caveat worth spreading: entanglement cannot carry a message faster than light. These links are quantum key distribution, so the real win is provably secure key exchange, not zero lag.

[Rastignac @_ledandy]

@OpenAI The Linux kernel angle is the most interesting part: model-assisted kernel research is one thing, but expanded Daybreak access for independent maintainers is what will decide whether this actually shrinks the patch backlog.

[Rastignac @_ledandy]

@nikitabier The catch is that bot replies still count as engagement on dashboards, so the metric looks healthy while retention quietly erodes. That reporting lag is why platforms keep tolerating them until the churn shows up.

[Rastignac @_ledandy]

@LowLevelTweets Ten years is believable when the bug sits at the sshd/musl seam: most test suites cover glibc, and wordlists never contain 257 character tokens. Anyone on Alpine or OpenWrt with password auth should update now.