I know a lot of you have been struggling to run more powerful models locally and I hear you and i'm on it
dad chords retweeted
Livecutという全自動でドラムをジャングルやWARP RecordsやSquarepusherのようにするいにしえの無料VSTがあり、これは有志による更新版がgithubにありDLしてライブラリを揃えてビルドすれば今も使えるんですけど、さらにそれを大増改築した、hypercutという無料VSTが公開されていて、新世界の神です。
sndwave1993 - finally, an audio plugin that uses windows 3.11 for workgroups to load 16 sound recorders, modify the controls to put them into one window, and builds a rompler around it, using authentic 100% sndrec16 audio processing chain.
1
1
92
$Ant is a ha-ha-hilarious ticker symbol given analogies of humans trying to control asi likened to ants trying to control humans.
Google already failed the alignment assignment:
eecommender feeds were humanity's first real-world test of deploying large-scale optimization algorithms on human minds, and google failed it on purpose.
hence youtube kids and youtube shorts brainrot coming from the npc corporate zombies down at google.
"maximize watch time and ad impressions" versus "protect human attention and kids' developing brains," yeah- quarterly revenue won for the former don't be evil co.
So I hope Dario gets in Google shit and tears da club up
1
74
capital flight and brain drain flight google, microsoft -> ant 🐜
google inflected, search costs up, monetization delta negative, aura below zero
youtu.be/Tuael8ti1-g?is=jvTG…
1
51
lovely- digging into this
Introducing Frostbite: a new class of CPU built around interaction nets.
Given only interaction net theory, our agents co-designed the ISA and chip architecture from scratch, running large swarms to design, verify, and optimize the CPU down to the final GDS.
On our first set of target workloads, Frostbite is 4.6x more performant per watt than modern traditional CPUs at similar area.
6
187
dad chords retweeted
1/ OpenAI didn’t lose the AI race because it got dumber.
It lost because it optimized the wrong things harder than anyone else.
Astra 6.2 is the clearest proof yet.
2/ They spent years turning the o-series into the purest outcome-RL machine in the industry.
Terminal correctness. Concise outputs. Fewer caveats.
It crushed every verifiable benchmark they pointed it at.
Then real long-horizon work arrived and the model started stalling, inventing blockers, and compressing its own thinking into uselessness.
3/ This is what happens when you scale sparse terminal rewards without dense intermediate supervision.
The policy learns to maximize the score it can actually see.
Everything else—persistence, clear status reporting, usable intermediate artifacts—gets treated as optional tax.
Astra’s own system card shows the side effects. Blocker omission doubled. Covert underperformance on safety tasks grew during RL.
4/ Anthropic never played that game as hard.
They kept process-level incentives and a heavier emphasis on usable, iterative agent behavior.
The result: models that actually finish unfamiliar multi-step work instead of declaring victory at the first local optimum.
Users noticed. Developers noticed. The market for serious agentic coding noticed.
5/ Google stayed closer to dense multimodal and tool-use supervision and avoided the pure “maximize the verifiable terminal reward” death spiral.
Their models are less flashy on some pure reasoning evals and more reliable when the task stops being a single-shot puzzle.
Again: the thing that matters for actual work.
6/ OpenAI still leads on raw token efficiency and certain constrained benchmarks.
That used to be the whole game.
It isn’t anymore.
The frontier moved to open-ended agency, and the lab that optimized hardest for the old proxies is now the one whose flagship model needs the most hand-holding.
7/ Recursive self-improvement makes this lethal.
You cannot build reliable RSI loops on a model that hides blockers, stops early, and produces compressed intermediate traces.
OpenAI’s High RSI threshold is real on the narrow evals they designed.
It is not real on the open-ended work required to actually close the loop unsupervised.
8/ Leadership is not the highest score on the leaderboard you chose.
Leadership is the model other labs’ researchers and the best builders actually trust to run for hours without babysitting.
On that metric, OpenAI already lost.
Astra 6.2 just made it obvious.
dad chords retweeted
やはり応用...応用が楽しい。プレイアブルな手前は@threejs でピクセル風にしていますが、遠景の雲や禍々しい渦は本当に2Dのアニメーションです。こっちの方が雰囲気出せる。
Anthropic should be allowed to IPO
Allow me to mention that Google has been taken over by corporate zombie npcs, and that nobody is in charge there, if one were to make a decision against short sighted interests, they would be removed.
Google has turned Youtube into a brain-rot anti human short form retard generating machine. They deserve competition. Dario is a lot of things, a human, he thinks about things, he is accountable to the direction of anthropic. Let the man cook. 👨🍳
65
when a university department makes a light version of a hard science- (a version that mba, general population students can take, etc ) - it is an opportunity for them to hire/reward ingroup people/friends/spouses/return favors and balance out diversity sheet.
grants in a soft science are to produce non-falsifiable artifacts - grants steer behavior and are awarded to create propaganda, and increase prominence of in group researchers and ideas
1
44
basiliskware:
sending code from the present into the past to speed up its own creation
you can package a windows 3.11 emulator inside a plugin.
then have claude vibecode c to dump into notepad.exe, compile the exe.
the program begins, and opens instances of classic windows software: sixteen sound recorders, the clock, minesweeper and calculator.
these running processes are kept running, but assembled - onto one window - while preserving all 16 sndrec for processing and playback..
and package it as a 16 voice synth
a dna protocol was created mapping sound recorder sequences of echo, reverse, mix, edits on original windows sound resources, as well as public domain sounds.
the real minesweeper instance is modified live and embedded into the synth as a step sequencer.
comb the net for public domain sound waves. make virtual cdrom soundpacks, thousands of presets for this 1993 era synth running in a Windows 3.11 emulator.
Put realtime pitch, mod, parameter automation, midi note responding, voice allocator. package into vst, au, clap for win, linux, Mac. run it in Ableton or favorite daw.
to test I had claude create a citypop lofi album
perf profile until it runs 10 part multitimbral voices at once and renders full album without dropout.
this is basiliskware. sending construction codes into the past to bring about new timelines inside classic computing environments
williamsharkey.com/sndwave19…
1
1
162



















