Das Ulises

Soy un astrocito pero la neta quería quedarme como CNS-progenitora. Chairo y moreno. Le hago a la complejidad y biología de sistemas. Complejo como un café.

[Das Ulises @uarayon]

"Y deletreas el nombre del Caos. Y no puedes dormir si no destapas el frasco de pastillas y si no tragas una en la que se condensa, químicamente pura, la ordenación del mundo." https://t.co/3sLuHy935q

[LaurieWired @lauriewired]

Physical Hardware is often (artificially) softlocked by firmware.

One of the biggest “unlock” hacks in history happened on IBM’s POWER8.

IBM Servers are kinda weird. They often ship with 40+ processor cores installed, but you only get to use whatever was licensed. As a business, you could then “upgrade in place” later, once you had the money.

Well, unless you take advantage of a funny firmware bug.

IBM POWER workloads run on a hypervisor, which divides physical system resources into pools.

Due to a synchronization bug (sort of like a double-spend in finance), if you moved pools around fast enough, you could add resources faster than they were removed!

The author went from 8 -> 40 POWER8 cores. Depending on the server you bought, you could have gone even further, from 8 -> 192 cores!

Probably the biggest unlock (in terms of “allowed” FLOPS to “unlocked” FLOPS) in history!

[Mario del Desierto @VicioNez]

Se nos pasó el aniversario de este maravillosos suceso pero aún así siempre vale la pena recordarlo

[Das Ulises @uarayon]

@Legendaryy @kimmonismus The og comparison was against GLM 5.2 (no 5.1)

[Andre Watson 🧬 @nanogenomic]

You can ask Claude Science to install LigandAI (pip install ligandai) and reference https://t.co/wPhPK1PyZx to generate nearly instant peptides and come up with target discovery :)

[Gemini Notebook @Gemini_Notebook]

Doom scrolling but make it educational 🤓

Introducing Short Video Overviews in NotebookLM! Turn your most complex sources into 60-second, vertical videos that deep dive into any concept.

Rolling out now to Google AI Ultra and Pro subscribers on mobile & web (free users soon!)

[Chubby♨️ @kimmonismus]

IBM just unveiled a sub-1 nanometer chip breakthrough.

That honestly wasnt on my bingo card.

Its new 0.7 nm / 7 angstrom technology uses a 3D "nanostack" transistor architecture to vertically stack and stagger transistors.

IBM says it can fit nearly 100 billion transistors onto a chip the size of a fingernail, almost 2x the density of its 2 nm chip from 2021.

• up to 50% more performance
• or 70% better energy efficiency
• 40% SRAM scaling for AI workloads

Important caveat: this is still research, not a chip shipping tomorrow. IBM says production could happen as early as the next 5 years.

[Omar Sanseviero @osanseviero]

We're hiring our first FDEs!

Being a FDE at Google DeepMind has a fun flavor. You will:
🤝 Act as a technical bridge between research teams & strategic partners
🛠️ Build benchmarks and evaluation tooling
🚀 Work at the frontier of AI

Join us 👉
https://t.co/3ybAhNZngz

[Unsloth AI @UnslothAI]

Google releases Gemma 4 QAT. ✨
You can now run Gemma 4 at 3x less memory with near original performance.

Quantization-Aware Training (QAT) makes it possible to run Gemma 4 26B-A4B on 16GB RAM.

GGUFs: https://t.co/wQgEocxUId
QAT Guide: https://t.co/Nsm1yeGEHx

[Maya Bechler-Speicher @mayabechlerspei]

🤔 Why do we still rely on the final layer of an LLM, when different layers encode different information? 🤔

In our new work, “Improving LLM Final Representations with Inter-Layer Geometry” (ICLR 2026 Workshop on Geometry-grounded Representation Learning and Generative Modeling) we show that actually, LLMs do not have one “best” layer.
We introduce the Cayley-Encoder: an efficient and effective geometric encoder that learns one strong representation from all layer representations of the LLM, without biasing the representation toward any specific layer.
While adding at most 0.1% learned parameters to the LLM, the Cayley-Encoder achieves large empirical gains over LoRA fine-tuning, final-layer representations, expensive attention-based aggregation, and methods that optimize specific layers for the task.

[Google Gemma @googlegemma]

AI video editors can't edit what isn't indexed. Learn how a developer used Gemma 4 31B locally on a 5-year-old laptop to process and index a year of raw, unlabeled video, making it fully searchable. A great look at building local-first tools.

[Noam Brown @polynoamial]

After AlphaGo, the skill of human Go players noticeably improved. I suspect we will see a similar pattern in math.

[Quanta Magazine @QuantaMagazine]

We experience geologic vibrations in our bodies thanks to the Pacinian corpuscle, a type of touch sensor. When the fingerlike protrusions on the neuron’s axon vibrate and stretch, the cell is activated, and we feel rumblings. https://t.co/FRQupbPXnU

[Pedro Miramontes @pmiravi]

Con un gran pesar en el corazón, les informo que hoy falleció Luis de la Peña. Grande entre los grandes maestros e investigadores de la UNAM. Hobre culto y consecuente con sus idea politicas de izquierda.
Gran amigo.

[SnoowBall_ @Snoow_Ball_06]

Los cuidadores del zoológico de Chapultepec llevan cuatro meses intentando atrapar a un gato naranja que se mete al recinto del tigre de Bengala a robarle la comida y el gato siempre gana. 😹🐾

Nadie sabe por dónde entra, nadie sabe a qué hora llega, pero la cámara del recinto lo graba tres o cuatro veces por semana apareciendo de la nada junto a la charola de carne cruda justo cuando el tigre se queda dormido después de comer la mitad. 😹

El gato pesa como cuatro kilos, el tigre pesa trescientos y el gato come de su plato como si le estuviera haciendo un favor.

La primera vez que el tigre despertó y lo vio se quedó mirándolo con una confusión que según los cuidadores nunca le habían visto en seis años, como si no pudiera procesar que algo de ese tamaño tuviera ese nivel de atrevimiento. 😹😹

Intentaron cerrar los huecos de la reja pero el naranja encontró otro, pusieron trampa con atún y el gato se comió el atún y siguió de largo al recinto, le cambiaron el horario de comida al tigre y el gato también cambió el suyo. 🐾💛

Un cuidador que lleva veinte años ahí dice que en todo ese tiempo nunca vio a un animal entrar voluntariamente al espacio de un depredador tres veces más grande y salir caminando tranquilo con la panza llena.

El tigre ya ni se levanta cuando lo ve, solo abre un ojo, lo mira comer, y vuelve a dormir como quien aceptó que hay batallas que no vale la pena pelear aunque las puedas ganar. 😹💛

El naranja sigue invicto y los cuidadores ya le pusieron nombre en el registro interno aunque oficialmente no existe. 😹

[PrismML @PrismML]

Today we’re announcing Ternary Bonsai: Top intelligence at 1.58 bits

Using ternary weights {-1, 0, +1}, we built a family of models that are 9x smaller than their 16-bit counterparts while outperforming most models in their respective parameter classes on standard benchmarks.

We’re open-sourcing the models under the Apache 2.0 license in three sizes: 8B (1.75 GB), 4B (0.86 GB), and 1.7B (0.37 GB).

[Qwen @Alibaba_Qwen]

🚀 Qwen Code v0.14.0 – v0.14.2 are now available

Channels：Control Qwen Code remotely from Telegram, DingTalk, or WeChat — send a message from your phone, get results on your server
Cron Jobs ：Schedule recurring AI tasks — auto-run tests every 30 min, pull & build every morning, monitor logs on a timer
Qwen3.6-Plus ：New flagship model with 1M token context, 1,000 free daily requests
Sub-agent Model Selection：Assign different models to sub-agents — use a powerful model for the main task, a fast one for subtasks, save tokens without sacrificing quality
/plan：Enter planning mode before execution — AI maps out all files and steps first, you confirm, then it executes
Follow-up Suggestions：AI suggests 2-3 next steps after completing a task — "Add unit tests?" "Check similar files?" — click to continue
Adaptive Output Tokens ：Default 8K output, auto-escalates to 64K when truncated — no more manually tuning max_tokens
Ctrl+O Verbosity Toggle ：Switch between verbose and compact output mid-conversation — debug mode when you need it, clean mode when you don't

📋 Full changelog: https://t.co/zPMyCVIskN

[Omar Sanseviero @osanseviero]

Introducing a Visual Guide to Gemma 4 👀

An in-depth, architectural deep dive of the Gemma 4 family of models. From Per-Layer Embeddings to the vision and audio encoders. Take a look!

[Prince Canuma @Prince_Canuma]

Gemma 4 31B running with TurboQuant KV cache on MLX 🔥

128K context:
→ KV Memory: 13.3 GB → 4.9 GB (63% reduction)
→ Peak Memory: 75.2 GB → 65.8 GB (-9.4 GB)
→ Quality preserved

TurboQuant compression scales with sequence length, so the longer the context, the bigger the savings!

Try it out:

> uv run mlx_vlm.generate –model google/gemma-4-31b-it –kv-bits 3.5 –kv-quant-scheme turboquant

Note: Decode speed drops (~1.5x) due to kernel launch overheads, we are aware and will fix in coming releases.

[Omar Sanseviero @osanseviero]

As part of the Gemma 4 release, we're launching Agent Skills: an Android app experience where you can import different skills and have Gemma 4 E2B reason and use the skills!

Running entirely in the phone, available in the Google PlayStore. Try it now!

[Not Even Wrong @notevenwrong]

New post: Langlands Program 2.0 https://t.co/xsNN2Ubf9n