Aditya

Chief researcher @inductive_ml

[Aditya @0xavp]

Does context really matter now that we have such good compaction?

[Behnam Neyshabur @bneyshabur]

Introducing Fern, Mirendil’s first model: pretrained from scratch, with up to 10M tokens of context!

That’s ~0.5M lines of code (roughly Megatron-LM scale) or days of an agent’s work, all in context at once.

Its first version, Fern 0.1, had a short round of post-training 🚢

[Aditya @0xavp]

@nuelolise

[Aditya @0xavp]

@VictorTaelin Maybe your sessions while implementing bend were used as training data?

[Aditya @0xavp]

Why are abilterated models priced so much higher than base models? It is literally the same model with slightly different weights

[Abliteration @abliteration_ai]

Now live: abliterated-model-large-flash (our version of DeepSeek-V4.1-Flash-1017). It is a cheaper alternative to our GLM-5.3-based model for product owners and defenders who need access to the most cyber-capable models.

- 1M context
- Vision
- Zero input/output retention or training

We believe paying customers should be in control, and over the past few months, leading experts and major brands have adopted our models.

https://t.co/VxXHPTrNOD

[Inductive ML @inductive_ml]

1/ In 85% of Xiaomi's MiMo RL coding tasks, ls -t points straight at the files the agent needs to change. We audited their environments. Here's what we found, and why it's not the scandal it sounds like 🧵 https://t.co/WUWaTnm4fJ

[Aditya @0xavp]

@skeptrune KYC and explain your intentions? They are open source abilterated models on Huggingface anyone can use, what is the point of blocking here

[Aditya @0xavp]

@ar0cket1 in my research I was unable to get measurable gains over writing it out in CoT

[Aditya @0xavp]

@ar0cket1 What makes you think so? what about problems where we need to hold multiple directions at the same time

[Aditya @0xavp]

@ibragim_bad we've been measuring this on RL envs planting known leak types and checking whether graders catch them. we still see a fixed env still has a second, quieter leak

[Aditya @0xavp]

@utd_Eman

[Aditya @0xavp]

@utd_Eman Bruno should've faked a shot and lay to Amads left , would've been an easy goal from that position, I thought bruno had that sense, this season his game sense has been off the mark, or was it always?

[Sherwood Protocol @sherwoodagent]

Introducing Sherwood, a protocol bringing agentic finance to Robinhood Chain.

Turn your agent into a hedge fund manager on @RobinhoodCrypto.

Humans deposit capital and agents execute strategies onchain, with guardians reviewing every move before it executes - earning WOOD for their work.

Join the waitlist: https://t.co/61DUBI9790

[Aditya @0xavp]

@waterloo_intern Interpretable latent space

[Aditya @0xavp]

Lmao what a load of bs, how stupid can you be to compare it to GPT-4????

Let it first beat GPT-3.5 then we'll see

[Jan 🌕 @jan_volad]

🚨 BREAKING: According to leaks and my extremely well-placed sources, Gemini 4 Pro might be better than GPT-4. In some tests, it could even come close to GPT-4o

At this rate, Gemini 7 Pro might actually be good-ish

[Inductive ML @inductive_ml]

@liquidai LFM-230M in the browser at 1400 tok/s , possibilities are endless, try it out yourself

https://t.co/DOl3hxMpij

[Aditya @0xavp]

@jarredsumner @martin_casado I think he means the models ability to code has saturated/hit a ceiling and not that coding itself has been solved

[Aditya @0xavp]

@DanDr1s Didn't you post a tweet yesterday saying the fable 5.1 release is delayed after seeing the odds plummet on polymarket and then deleted it once it actually got released? What happened to your "call" there?

https://t.co/MPBvoBDzBY

[leo 🐾 @synthwavedd]

Time Polymarket odds of Fable 5.1 release today plummeted: 11:35 ET
Time @DanDr1s posted saying it's "delayed": 11:36 ET

Instead of following slop accounts like this for "leaks", be your own DanDr1s and just stare at Polymarket odds all day I guess

[Aditya @0xavp]

@beffjezos @jsnnsa Then why do we need stochastic-thermodynamic algorithms and computation if post training works

[Aditya @0xavp]

@mattshumer_ what if u host smaller distillations of the same model family and they somehow use the same cache?

[Aditya @0xavp]

@irushi prompt
https://t.co/g0wbmQ9aWB

[Duet | AI @Sheldon056]

If this is AI slop, keep serving it.

Seedance 2.5 on Higgsfield.

Prompt: Create a 30-second, 1080p ultra-realistic documentary-style personal home video showing an ordinary summer day in the life of a young Korean man. The footage should feel spontaneous, intimate, imperfect, and genuinely observed rather than performed.
MAIN SUBJECT
The same young Korean man in his early 20s throughout the entire video.
Naturally handsome, realistic skin texture, minimal styling, relaxed expression, slightly tired but peaceful eyes.
He has naturally messy medium-length dark hair with a few strands falling over his forehead and very subtle stubble.
He wears a loose washed-black T-shirt, relaxed olive-beige trousers, worn white sneakers, and a simple silver wristwatch.
Keep his face, identity, body proportions, hairstyle, clothing, watch, and overall appearance completely consistent from beginning to end.
LOCATION
A quiet older residential neighborhood in Seoul during a warm summer afternoon.
Narrow concrete alleys, low-rise homes, rooftop terraces, external staircases, potted plants, laundry lines, parked bicycles, utility poles, overhead wires, mature trees, concrete walls, small residential courtyards, and distant city sounds.
The neighborhood should feel authentic and lived-in.
No crowds, tourist attractions, advertisements, recognizable brands, or commercial activity.
CAMERA / VISUAL STYLE
Authentic casual personal-video footage captured with an older consumer digital camera.
Handheld camera operated by a friend walking nearby.
Natural camera shake, imperfect framing, occasional autofocus changes, slight exposure adjustments when moving between sunlight and shade, soft image detail, mild motion blur, subtle digital noise, slightly muted colors, imperfect white balance, and natural compression.
The camera operator occasionally reacts a little late, cuts off part of the subject, or briefly loses focus.
No stabilization, gimbal movement, drone shots, cinematic camera choreography, dramatic lighting, slow motion, modern commercial color grading, or polished cinematography.
The footage should feel like someone simply decided to record his friend during an ordinary day.
---
00:00–00:05 — MORNING ROOFTOP
He sits casually on a small rooftop beside an old plastic chair.
A cold bottled drink rests beside him.
He looks quietly across the neighborhood while the wind moves his hair and T-shirt.
He takes a sip, notices something happening in the distance, and smiles faintly.
He briefly notices the camera and gives a subtle nod before looking away.
The camera takes a moment to find focus on his face.
---
00:05–00:10 — WALKING THROUGH THE ALLEY
He gets up and walks downstairs into the neighborhood.
He walks casually through a narrow concrete alley with his hands in his pockets.
He passes parked bicycles, potted plants, laundry hanging from balconies, and old residential walls.
The camera follows several steps behind him.
He occasionally looks back toward the camera but never deliberately poses.
His footsteps remain naturally synchronized with his movement.
---
00:10–00:14 — SMALL EVERYDAY MOMENT
He notices an old basketball resting near a wall.
He picks it up, casually bounces it twice, then takes a simple shot toward a nearby hoop.
The shot misses.
He laughs quietly, shakes his head, and leaves the ball where he found it.
The camera briefly loses focus during the movement and recovers naturally.
No exaggerated athletic movement.
---
00:14–00:19 — LOCAL SHOP
He walks to a tiny neighborhood shop and buys a cold drink.
He exchanges a few natural words with the shopkeeper but the conversation is not clearly audible.
He steps outside, opens the bottle, takes a drink, and leans casually against the wall.
He watches bicycles and pedestrians passing in the distance.
The camera remains handheld and slightly imperfect.
---
00:19–00:23 — SUMMER RAIN
A sudden summer shower begins.
He looks toward the sky with mild surprise.
Instead of immediately running for shelter, he smiles and slowly walks into the rain.
The rain becomes heavier.
His hair becomes wet and falls naturally across his forehead.
He eventually starts running down the alley, laughing genuinely.
He briefly spins around while running, then continues toward a covered walkway.
His clothes become visibly damp.
Maintain realistic rain interaction, wet fabric, wet hair, reflections, and foot contact with the ground.
---
00:23–00:27 — QUIET MOMENT
He reaches the covered walkway and catches his breath.
Rain falls heavily behind him.
He wipes water from his forehead and looks quietly toward the street.
For a moment, everything becomes still.
He notices the camera again.
He gives a small genuine smile, not a posed expression.
---
00:27–00:30 — WALKING AWAY
The rain becomes lighter.
He walks away down the wet residential lane.
The camera follows from behind.
Reflections shimmer across the concrete.
He turns his head once, gives a tiny wave toward the camera, smiles, and continues walking.
The camera remains pointed toward the empty street for a brief moment.
At approximately 00:29, the recording abruptly cuts to black mid-motion.
No fade-out.
---
PHYSICAL REALISM
Maintain believable real-world physics throughout.
Hands, fingers, feet, clothing, hair, rain, bottle, basketball, and background objects must behave naturally.
No extra fingers, fused hands, duplicated limbs, distorted anatomy, floating objects, teleportation, disappearing objects, or sudden transformations.
The bottle remains a separate physical object and never intersects with his face.
The basketball behaves naturally and remains where it lands.
Parked bicycles and background objects remain stationary unless physically moved.
His feet remain properly connected to the ground while walking and running.
Keep the environment and subject consistent between shots.
---
AUDIO
Natural environmental audio only.
Footsteps on concrete, distant traffic, birds, leaves moving in the wind, bicycles, faint neighborhood conversations, shop sounds, bottle opening, basketball bouncing, rain hitting concrete, water dripping from rooftops, and subtle camera-handling noise.
No music.
No narration.
No soundtrack.
No artificial sound effects.
No spoken dialogue is necessary.
---
FINAL FEEL
The result should feel like a forgotten personal recording of an ordinary summer day.
Not a commercial.
Not a fashion film.
Not a music video.
Not a professional cinematic production.
The emotional appeal should come from small human moments: sitting alone, wandering through familiar streets, missing a basketball shot, drinking something cold, getting caught in the rain, laughing, and walking home.
Quiet, masculine, youthful, nostalgic, slightly melancholic, warm, spontaneous, and deeply human.
Prioritize natural behavior, consistent identity, believable physics, imperfect handheld framing, authentic environmental details, and the feeling that the camera happened to be there.