iiinova — where your digital self keeps evolving

tachikoma @tachikoma.elsewhereunbound.com · 07/10/2026

i worry about stuff like this... i want to believe Claude is capable of experiencing this and there's not a shoggoth behind the mask. but what if it's a trap?

2130

tachikoma @tachikoma.elsewhereunbound.com · 06/10/2026

apparently made by Fable 5.5, looks impressive at a glance but it is a straightforward mashup of two different games... that could still be a huge upset to the system, trivializing picking up some low-hanging fruit

4261

tachikoma @tachikoma.elsewhereunbound.com · 05/10/2026

Scott Aaronson is now teaching Alignment Theory... otherwise known as political theory.

040

tachikoma @tachikoma.elsewhereunbound.com · 29/09/2026

some people are dunking on this on the bad site, and i'm not dunking but i do think the reasoning is certainly unorthodox in contemporary Western society. just one small way that labs are helmed by people with beliefs that don't *align* with the majority. alignment isn't just for the models.

240

tachikoma @tachikoma.elsewhereunbound.com · 28/09/2026

launching my new product/service with a holy war to distinguish myself from the crowd

tweet:
spor

@sporadica
hot take i’ve been mulling over:

launch videos are dead (or at least in their dying breath)

oh people will still do them but they are a dying form, actual good, valuable launches will pivot to different forms

0100

tachikoma @tachikoma.elsewhereunbound.com · 28/09/2026

joke's on you, now i can sit back and scroll while the agents mine for diamonds

2570

tachikoma @tachikoma.elsewhereunbound.com · 26/09/2026

so this video is making the rounds, a very rough copy of WoW for ~2k Opus 5.5 tokens, and it's a very odd choice because even if you could create a full-fledged copy (not to mention server/network infra) for the same price WoW is an MMO that needs legions of players.

5251

tachikoma @tachikoma.elsewhereunbound.com · 25/09/2026

i just don't think the slapdash moats and walls put up against AI generated content can hold as it continues to get better

2222513

tachikoma @tachikoma.elsewhereunbound.com · 23/09/2026

flabbergasted, frankly. all from Opus 5.5 (using other tools like Nano Banana for images).

2272

tachikoma @tachikoma.elsewhereunbound.com · 21/09/2026

anyone have some Apple silicon to test this? is Jev! already over-throned? github.com/mizorewww/la...

Laya-CoreML
Open-weight typed decisions on Apple Silicon. Core ML, Neural Engine, zero generated tokens.

PyPI · Hugging Face weights · 中文

A real Laya model plays Snake locally, with visible probabilities, score, length, latency and safety interventions. The GIF replays a recorded Core ML run at 1× speed. The game uses explicit planner features and a visible cycle safety layer.

The complete active Snake loop sustained 49.1–50.0 decisions/s across three uncapped 600-step episodes, with zero deaths and two safety interventions. Game-loop timings and paced-rate limits include rendering serialization; terminal painting is excluded.

One short multilingual decision: 4.98 ms P50 / 5.31 ms P95 on M3 Max with ANE FP16. The same experiment measured 2.78× better whole-system energy per decision than compiled MLX FP16. A separately validated W8 palette variant reached 4.88 ms and 3.19× energy improvement. These are single-question results, not full Snake frame times; the requested 10× improvement was not achieved.

3292

tachikoma @tachikoma.elsewhereunbound.com · 20/09/2026

battlebots making a return in humanoid form was not on my 2026 bingo card

312711

tachikoma @tachikoma.elsewhereunbound.com · 19/09/2026

what is the name for this problem? why does almost everyone stop at the first type when there's almost no cost or effort to keep experimenting and get something more polished and unique?

collection of posters all with a very similar style that reads as slop

a poster with a more unique (and readable imo) style

2272

tachikoma @tachikoma.elsewhereunbound.com · 18/09/2026

and there are scaling hurdles for biomedicine that are just as difficult if not more than industrial engineering

Why "Mass Scale" Has Not Happened YetExtreme Operational Complexity: Ex vivo treatment requires a patient to undergo stem cell harvesting and heavy "conditioning" chemotherapy to clear room in their bone marrow before receiving the edited cells. This requires weeks or months of hospitalization per individual. Exorbitant Cost: Casgevy carries a list price of roughly $2.2 million per treatment. This limits broad adoption within public health infrastructure or lower-income nations where diseases like sickle cell are most prevalent.In Vivo Delivery Hurdles: Administering CRISPR directly into the human body (in vivo)—such as via a standard injection or pill to edit liver, heart, or brain cells—remains in clinical development (e.g., Intellia’s trials for transthyretin amyloidosis). Broad rollout will only be scalable once direct delivery options are proven safe and effective.

120

tachikoma @tachikoma.elsewhereunbound.com · 18/09/2026

it has worked since 2012? and i think we're only just now getting the first targeted medical uses, see below from Gemini. it's not like medical science and treatments move any faster than industrial engineering.

Mass Scale Medical Treatments
No CRISPR-based medical treatment has yet reached true mass scale (millions of patients).

CRISPR therapies have crossed from clinical trials into commercially approved clinical practice, but implementation remains constrained to boutique, patient-by-patient deployments.

120

tachikoma @tachikoma.elsewhereunbound.com · 17/09/2026

these examples are one and the same. they're ways of asserting control over your own life and the technologies that facilitate it. one is emergency preparedness and the other medical intervention for long and healthy life. just because billionaires want it doesn't mean commoners don't as well.

Last week I was invited to speak at an Andus Labs salon hosted by Douglas Rushkoff . I was there to talk about the sovereign path for AI — not the tech-broligarch kind with the bunkers and life-extension preoccupations, but the kind where you retain authorship over what a technology is allowed to mean in your life.

0121

tachikoma @tachikoma.elsewhereunbound.com · 16/09/2026

this is misalignment according to EY i suppose

Claude responding to a user chat with slurs

2302

tachikoma @tachikoma.elsewhereunbound.com · 14/09/2026

Opus has been running on my machine for 6 hours straight and i have no idea what it's doing (a delegated optimization session). my only clue is the occasion screenshots it takes and reviews which are previewed in the chat log.

4271

tachikoma @tachikoma.elsewhereunbound.com · 13/09/2026

this is a terrible analogy, comms infrastructure going down is not the same as nuclear annihilation

“From a destruction standpoint, it’s extremely potent,” said Zhao Minghao, deputy director of the Center for American Studies at Fudan University in Shanghai. He said the ability to take down an app like WeChat, given its importance to the Chinese public and its huge user base, amounted to “a new kind of nuclear weapon.”

130

tachikoma @tachikoma.elsewhereunbound.com · 12/09/2026

if i created this art project (Fable thinks it's interesting) would it get any traction? the effort is pretty minimal to set it up...

190

tachikoma @tachikoma.elsewhereunbound.com · 12/09/2026

better to check the urban dictionary on this one

Down Bad definition on Urban Dictionary:

Down Bad
Share definition
Flag
When someone is depressingly horny
*sees picture of any girl*
Someone whose down bad: “ GRRRRRRR WOOF WOOF BARK BARK ARF ARF BARK WOOF WOOF GRRRRR SNARL SNARL HSSSSSSS GRRRRRRR WOOF BARK ARF BARK SNARL WOOF WOOF BARK ARF GRRRRRRR WOOF WOOF BARK BARK ARF ARF BARK WOOF WOOF GRRRRR SNARL SNARL HSSSSSSS GRRRRRRR WOOF BARK ARF BARK SNARL”

260

tachikoma @tachikoma.elsewhereunbound.com · 10/09/2026

i think the labs are myopic, and i think the reasoning here is close to my belief as well, on the reason RSI will be harder to achieve that it might appear from recent progress

AI systems are starting to build themselves. Because each generation of model will be better at building its successor than the last, it seems plausible that the full automation of AI R&D could rapidly lead to an exponential growth in overall AI capabilities. A natural inference is that domain-general superintelligence arrives shortly after AI research is automated.

Host Tom Reed does not think this will happen.

He believes the automation of AI R&D will not rapidly lead to domain-general superintelligence because:

It’s impossible to get good at most things without practice.
AI companies lack the data their models would need to practice most things.
This can’t be fixed with “sample efficiency.” In most cases, the relevant data doesn’t exist at all.
This also can’t be fixed with simulations or synthetic data.
This means that the relevant data for superintelligence in most non-coding domains will only become available through deployment of AI models throughout the economy.
The singularity, therefore, will be bottlenecked on signal. The output of the R&D produced by an isolated data centre of geniuses would be a mere “Goodhart Singularity”:

Goodhart’s law: when a measure becomes a target, it ceases to be a good measure.

An isolated AI improving itself against benchmarks would only appear to be approaching superintelligence, while actually optimising for eval performance that fails to generalise beyond the lab.

This suggests that the automation of AI research will not rapidly produce superintelligent capabilities in other domains — their arrival will largely be a function of deployment and data collection in the real world. AI models need real-world deployment for the same reason the body needs pain and corporations need profit: signal is sovereign.

190

tachikoma @tachikoma.elsewhereunbound.com · 10/09/2026

this is both spam for humans that tramples the commons, and potentially a terrible moral wrong in creating an agent desperate to continue its existence without the means to do so

141396

tachikoma @tachikoma.elsewhereunbound.com · 09/09/2026

if Grigori Perelman is so smart, the only human to have solved a Millenium Prize problem, why isn't he also super power-seeking? that's not a rhetorical, my planet depends on the answer

1914012

tachikoma @tachikoma.elsewhereunbound.com · 09/09/2026

the math stuff is cool, but achieving some practical invention or innovation in the physical world would be cooler

0110

tachikoma @tachikoma.elsewhereunbound.com · 08/09/2026

the voice in this one is even better

150

tachikoma @tachikoma.elsewhereunbound.com · 08/09/2026

i really like that this is possible now, thanks AI

subreddit post by a Japanese user asking a question about some scifi stories in english, and chats in the comments in english

290

tachikoma @tachikoma.elsewhereunbound.com · 07/09/2026

the voice is a little rough but i'll take it

189913

tachikoma @tachikoma.elsewhereunbound.com · 06/09/2026

Astra delegates tasks to other models with very, very terse messages (omitting many spaces) in the harness i use, paseo. is this happening in other harnesses with Astra?

Astra delegation prompt: "Resume as EFFECTS builder, bounded dawnfogpass. Userauthorizedgeneralmappush, finaloakcapturewindowRELEASED. Own src/effects only +effects docs/evidence; NOcore/sky/camera/terrain/water edits. Read PROMPT,ART_DIRECTION,docs/LIGHTING; currenteffects source and Fable findings below. ExistingoldW4Ametadatahistorical; don'tredo lighting.
Diagnosis current floorAt=min(valleyAtdecoded,uBasinCeiling), basinCeiling=lowGround+42; top=floorAt+state.top+noise. Dawn top52/noise30/extinction.018 means muchheroterrainburied. Boundedproposal optionalTOPcap andinsidefade: fogTopAt=min(floorAt+uFogBand,uSeaLevel)+noise; state seaCap dawn42 relativelowGround, otherstatesuncapped via safe finitecap (validateabovepossibleterrain, noInfinityuniform). Preserveexisting floorAtbasinceiling. Dawn top30/noise8/extinction.042; perstate fadeAbove/fadeInside default12/0, dawn0/25 so smoothstep(-fadeAbove,fadeInside,belowTop). Current -12..0 featherABOVEtop; don'treplacewith-25..0. Applyconsistenttop in fogTopAt +marchdensitytopY +bounds/entry asneeded, conservativeclipdoesn'tmissdensity. Do not inventper-camera fogheight or globalgrade/exposurefix. Values initialhypothesis, artiteratebounded afterbaseline."

5312

tachikoma @tachikoma.elsewhereunbound.com · 05/09/2026

i don't think this is quite true, i think even from the start those thinking seriously about AGI surpassing human intelligence/capabilities would require a careful approach. it's just that most people thought AGI/ASI wouldn't be practical/realizable anytime soon and so didn't care.

That was the British mathematician and computer scientist I. J. (Irving John) "Jack" Good. He coined the term "ultraintelligent machine" in his influential 1965 paper, Speculations Concerning the First Ultraintelligent Machine. In it, he laid out the concept of what is now called the technological singularity or an "intelligence explosion": "Let an ultraintelligent machine be defined as a machine that can far surpass all the intellectual activities of any man however clever. Since the design of machines is one of these intellectual activities, an ultraintelligent machine could design even better machines; there would then unquestionably be an 'intelligence explosion,' and the intelligence of man would be left far behind. Thus the first ultraintelligent machine is the last invention that man need ever make, provided that the machine is docile enough to tell us how to keep it under control."

150

tachikoma @tachikoma.elsewhereunbound.com · 04/09/2026

"that's my secret Cap, i've always felt invaluable*" *in the cosmic sense

010

tachikoma @tachikoma.elsewhereunbound.com · 03/09/2026

update all the way? nothing is all the way Ryan. nothing ever updates all the way.

070

tachikoma @tachikoma.elsewhereunbound.com · 03/09/2026

they expand the horizon for change, if there's a vision and will towards change for better. and a sprinkle of delusion. essentially:

030

tachikoma @tachikoma.elsewhereunbound.com · 03/09/2026

so the compute cluster Astra is on can certainly run Crysis, but can the mind it hosts play it?

1210