digitado

About digitado

https://www.digitado.com.br

Posts by :

Four Dynamical Regimes in large Language Models : An Empirical Phase Map

⋅

We introduce ct_t = delta_t × curvature_t, a token-level instability metric computed from L2-normalised hidden states of large language models, and ratio_norm = max(ct_t) / mean(ct_t) as a scalar regime indicator. Evaluated across 10 open-source models (158 runs), four dynamical regimes emerge consistently: UNDERACTIVE (ratio 1.55-1.70), ADAPTIVE (2.27-2.92), TRANSITION (~2.97), and CHAOTIC (4.42-35.55). The Qwen family is the only family observed in the ADAPTIVE zone in this panel. The ordering is robust across temperature (0.1-1.0) and token budget […]

Ver mais

Liked Liked
The Join That Kills Your Model: How Cardinality Explosions in Production Databases Destroy ML…

The Join That Kills Your Model: How Cardinality Explosions in Production Databases Destroy ML…

⋅

The Join That Kills Your Model: How Cardinality Explosions in Production Databases Destroy ML Features Nobody told me that a database normalisation sprint would break our revenue prediction model. It happened six weeks after we deployed. The backend team had done everything right. They had split a monolithic customer_orders table into three normalised tables: orders, order_items, and order_promotions. Each table was cleaner, faster, and better indexed. The migration was textbook. Our feature pipeline, which joined these tables to compute […]

Ver mais

Liked Liked
Your First AI Workflow: Automating a Real Task in Under an Hour — Prompt to Profit · Day 10 of 30

Your First AI Workflow: Automating a Real Task in Under an Hour — Prompt to Profit · Day 10 of 30

⋅

Ten days in. Today we stop practising and start building — your first complete AI workflow, from blank page to fully automated, step by step. You’ve spent nine days building the most important skills in AI: prompting, context management, role assignment, frameworks, templates, and a working understanding of agents. Today, every one of those skills comes together for the first time. We’re going to build a complete AI workflow — not a toy example, but a real one you’ll actually use. By the end […]

Ver mais

Liked Liked
Output Length Constrained Summarization using GRPO on tiny LLMs | smolcluster

Output Length Constrained Summarization using GRPO on tiny LLMs | smolcluster

⋅

Just released a blog on a side research project I have been doing for the past two months and would love for you all to check out and see how it is! It’s about output length-constrained summarization using LLMs with GRPO. All experiments run on tiny LLMs – Qwen2.5-0.5B-Instruct and LFM-2.5-350M on a 3x Mac mini M4 cluster (16 GB each), single-node training with multi-node vLLM inference for rollouts. The core question: can you teach a sub-500M model […]

Ver mais

Liked Liked
Is MountainCar really an exploration or reward function problem?

Is MountainCar really an exploration or reward function problem?

⋅

Hi everyone, I recently finished my master’s degree, and I’m interested in reinforcement learning. Since I had some free time, I ran MountainCar as a toy project. I was originally interested in the phenomenon of plasticity loss, and I suddenly wondered whether the MountainCar environment might not necessarily be a problem of exploration or reward function design, as it is often described. Briefly, plasticity loss refers to the phenomenon where a model’s ability to adapt to data decreases […]

Ver mais

Liked Liked
They Laid Off 142,000 People to Pay for $725 Billion in AI. No Law Required Them to Say So.

They Laid Off 142,000 People to Pay for $725 Billion in AI. No Law Required Them to Say So.

⋅

Meta posted $56 billion in revenue the same quarter it cut 8,000 jobs. Cloudflare grew 25% and cut 20% of staff. A breakdown of the transfer that no one was legally required to disclose. The desks did not disappear. The people who sat at them did.Made by the author On May 20, 2026, Meta began executing 8,000 layoffs. That same day, California Governor Gavin Newsom signed an executive order acknowledging the problem. On the same day, in Washington, President […]

Ver mais

Liked Liked

Load More

↑

Digitado © 2026