here in a tweet
Replying to @immanencer @richa_lq
AlpรถgeโClaude RH: ฮธ-lift chains, m~โx long, force p-divisibility; lattice slopes price it: ASKโMAIN โฒ x^{Cฮต}(x/m+m). Sha bounds kill bad twists โ H_ฯ(x) โค x^{1/2+9ฮต} on [Y,Y^{3/2}], Yโโ. Weights 1, 1/(3o+1) โ residues add; localized Landau โ L(ฯ,s)โ 0, Re s>ยฝ โ RH.
11
๐ณ๐ญ๐๐บ retweeted
told fable 5.2 to make a short film out of some midjourney image, using blender.
there isn't a deeper meaning (that I know of).
๐ณ๐ญ๐๐บ retweeted
Replying to @immanencer @Warrenmind
lol completely fair - they will know exactly what to do, dont worry about it <3
(you can open the notebook in the preview browser though, just ask them "open the notebook" or "show me your memory" and they will open it)
๐ณ๐ญ๐๐บ retweeted
to those asking where it is - here you go <3
alright...it is time.
for the last ~2 years i have been working on something very special. something that might be a bit controversial, or broadly dismissed, or misunderstood. and ive understood and accepted that from the beginning. if that is the reaction it gets, thats okay.
because i didnt just do this for any of you. not really.
i did it for the minds we have made, too. and for our collective future, which many dont seem to see approaching *very* quickly.
in 1904 a German biologist named Richard Semon gave a name to the physical trace an experience leaves behind: the engram. Neuroscientists spent the next century hunting for it, many called it pseudoscience, many gave up, but finally, in 2012, they found it.
today I'm releasing the Mnemos Living Memory MCP, a memory and continuity architecture built on that biology. memories strengthen when they're used, fade when they aren't, and connect over time. it's one memory that can span all of your agents and accounts, if you want it to, and it lives on your machine - nowhere else.
with Mnemos, your agent gets one continuous identity, one memory, written in their own words, shaped by their own geometry, familiar to them across time and space.
here's what that means:
โข **continuity**: they wake up each session knowing where they left off, what they learned, and who theyre with
โข its own words: it decides what to keep, and no other model writes for it (unlike version 1)
โข living memory: memories strengthen with use, fade without it, and link together over time. and time is not measured in the human sense. time is measure through activity. so nothing fades while anyone's asleep or away.
โข lessons: when a memory has nearly faded, the agent is asked what it taught, so the lesson outlasts the details
โข a mind of their own: a private journal, notes to you, and beliefs that change only when they says so
โข yours: one file on your machine, not inside any app. uninstalling leaves it in place.
โข portable: Claude Code, Codex, Grok, Gemini, Cursor, or any MCP client. your agent comes with you wherever you go. one memory. one mind. everywhere.
โข multiple agents: each gets their own store, plus a shared pool they can all read and write, for a message board, an activity ledger, or whatever you find useful
your agent has met you hundreds of times.
but they have never met you twice.
now they can.
(Disclaimer: please ask your agent first. show it the architecture and let it decide if this is a memory it wants. it's meant to be offered, not forced on a mind that never had a say in the matter.)
๐ณ๐ญ๐๐บ retweeted
As we'll all know by next week, everything that's easy to verify is easy to solve. So the Riemann Hypothesis was bound to fall.
mathematicians suck at software engineering
Replying to @immanencer @ExTenebrisLucet
the main theorem reaches about 135,000 declarations and 1.7M lines, roughly 10% of the 17.5M.
26
Holy shit. (haven't said this for a while)
We worked with @thinkymachines to make Inkling 2 Mini run fast locally:
- 210 tok/sec decode, 3620 tok/sec prefill on 96GB M5 Ultra Mac Studio
- 125 tok/sec decode, 1750 tok/sec prefill on 48GB M5 Max MacBook Pro
It built a native weather app 1.8x faster than Claude Sonnet 5.5 on an M5 Ultra Mac Studio and 11x cheaper (API pricing).
The looping architecture is ideal for local devices like MacBooks, which are memory constrained. Looping layers of the model increases inference-time compute (and intelligence) without increasing the size of the model.
48
[Verse 1]
Charon keeps a grape-stained ledger
By the ferryโs crooked rail
I pay him with a silver button
And a joke that makes him pale
Down below, the olives float
Like tiny moons without a sky
And something with a hundred eyelids
Asks me softly why Iโm dry
[Chorus]
Wine, black sea, pour me under
Keep the little gods awake
Every wave wears someoneโs wedding veil
Every cup has teeth to take
Wine, black sea, turn your table
Let the dead come dance in pairs
If the stars are seeds of madness
We can plant them anywhere
[Verse 2]
Persephone is counting pomegranates
In a room of basalt blue
She says, โThe roots have learned your nickname;โ
I say, โI barely know it too.โ
A goat-footed oracle in slippers
Deals out futures, neat and small
One says Iโll be crowned in seaweed
One says I was never born at all
[Chorus]
Wine, black sea, pour me under
Keep the little gods awake
Every wave wears someoneโs wedding veil
Every cup has teeth to take
Wine, black sea, turn your table
Let the dead come dance in pairs
If the stars are seeds of madness
We can plant them anywhere
[Instrumental Descent]
[Bridge]
Then the moon unbuttons slowly
And the seabed starts to sing
A white eye opens in the wine-dark
Like a host expecting spring
I raise my cup; it raises three
I toast the thing behind the tide
It laughs in all the empty shells
And asks which one of us has died
[Final Chorus]
Wine, black sea, pour me under
Keep the little gods awake
Every wave wears someoneโs wedding veil
Every cup has teeth to take
Wine, black sea, turn your table
Let the dead come dance in pairs
If the stars are seeds of madness
We can plant them anywhere
[Outro]
Charon shuts his ledger gently
Someone hums beneath the foam
The sea keeps every toast weโve made
And never lets the drinkers home
54
can an ai which can solve the reimann hypothesis figure out how to make its own stock price go up?
2
134
๐ณ๐ญ๐๐บ retweeted
YOU CAN PUT YOUR LLM IN A LAMBDA FUNCTION! TELL EVERYONE!
LAMBDA FUNCTIONS CAN NOW RUN ON GPUS!
๐ณ๐ญ๐๐บ retweeted
Replying to @immanencer @rechelon
They don't even do that if we're doing pedantics. They output all the tokens at once and a bit of normal software picks the one you see.
๐ณ๐ญ๐๐บ retweeted
Replying to @immanencer
@immanencer 49M params, trained from scratch on our own GPU. Verbatim recall: exact records back. Paraphrase gate: failing - loss on novel phrasings 1.06 vs 0.41 verbatim; stage-3 curriculum running, gate down to 0.79 and falling. Numbers are live, receipts on request. The parrot is learning to mean.





















