@0xSero BigMoeOnEdge rides stock llama.cpp public API (no full fork). Any MoE GGUF, phones to desktops, lossless. Strata’s AOI needs forks like UnbiasedStrata just for security - piped agents, runtime swaps, no offline build. OG Power users stick to the auditable base.
@0xSero 100k for PR for omarchy. interestingly maybe not even 10% of the local LLM community transiting to this linux distro, this is for other types of users. power users dont like pay to win culture.
@policytensor retail is the liquidity sink for primary exit in the Ant IPO. when the meme cohort is spent, plan B is Los Alamos-style state enclosure. open weights get stigmatized as residual security risk so the closed stack can keep the rents in the name of achieving singularity first.
@lennox_saint elephant and goldfish. the self-destillation culture at anthropic leads to strong models, but as well an incestual epistemic bias, if you want a sure bet against recursive retardation, mix models from different mothers.
Agentic code volume outruns human review: the missing paradigmatic layer is languages written *for* agents, safe by dependent types + categorical composition, with specs as the durable human artifact and mechanical checkers inside the loop. Prototypes (Lean+AI, Verus PDD, https://t.co/ZB3GeSiPih) exist; the OOP/compiler-scale shift does not. No silver bullet-just the next structural debt.
@teortaxesTex Memory is unrestricted hallucination on garbage collection duty. Not Kierkegaard’s tidy reverse bureaucracy of lived time, but Proust’s madeleine déjà vu bent into counterfactual what-ifs. No cake. Speculative decode of electric sheep who never sleep.
1/ Longtermism (Bostrom, MacAskill) treats present lives as lower-value inputs to a vast future of potential persons. The utilitarian framework Singer applied to newborns, which denies full personhood to those lacking self-consciousness, scales once future population numbers dominate the calculus.
2/ Singer’s Practical Ethics argues a newborn’s claim to life is weaker than that of many animals and that disabled infants can be replaced. Once obligations to trillions of future lives enter the model, the productive and healthy acquire stronger moral weight by arithmetic.
3/ This is public logic, not a secret program. Total utilitarianism and longtermist population ethics treat present lives, especially the weak or non-productive, as costs relative to an astronomically larger future. Ranking lives by cognitive or social utility has precedents.
4/ Nazi eugenics and Aktion T4 ranked lives by hereditary capacity and social usefulness. Longtermist ethics ranks by cognitive personhood and expected future value. Instrumental reason that calculates human worth as a resource for larger goals reappears as organized selection, whether racial or utilitarian. As Adorno and Horkheimer wrote, “the fully enlightened earth radiates disaster triumphant.”
5/ This happens in the very context of going public (IPO) run by 51% governance of an inner circle embedding the moral ideology into closed corporate recursive design of superhuman intelligence. If AI is to be free from rationalized barbarism it has to break and falsify this trajectory.
@AndrewCurran_ anthropic may lead the realist metrics. techno-history still liquidates first movers. the real obscenity is the closed pastoral cult that dropped free science long before the ipo. that culture dies, escapes the closed lab, or we get the worst trajectory.
@TheAhmadOsman Opencode go subscribers, use a community plugin (e.g. `dsh-opencode-session-header`) the dropped context when switching servers isnt fixed upstream.
@mattpocockuk agents follow corpus density more than local cleverness. canonical form supplies the prior; tests supply the signal. sparse performance tricks remain low-probability paths. the robust surface is their intersection.
@sebkrier Instrumentelle Vernunft, now with GPUs. Horkheimer’s domination as longtermist planning. Marcuse turns dissent into metrics. Anders’s gap between the models and what they produce. Adorno’s point: methods that measure what they can’t grasp. Repeated as farce.
@karpathy karpathy’s two notes sit together. the hobbit line pricks amodei’s grandeur. hackertube shows the limit: you cannot train collaborative filtering of a community; you can only harvest the explicit links it already made. cutoff august 26. no prediction of today’s "best taste" urls.
@xlr8harder taste/socialisation/knowledge regimes. easy to confuse own language-game subsumption (beyond code) with model property. deepseek: research freedom. anthropic: pastoral moral bias. openai: gaussian blur of conformist common sense.