Articles de praticiens IA (Medium)
Pas encore d'avis sur cette source.
Vous suivez Towards AI ? Donnez-lui une note et partagez votre avis.
An overview of AccessEval framework Ableism — discrimination or prejudice against individuals with disabilities. Source : Photo by Zhuo Cheng you on Unsplash Are LLMs more supportive and helpful when a prompt is made in a disability context, or do they get more negative and stereotypical? Does scaling the models automatically improve fairness in a model? How can we even measure such biases? These are some of the questions addressed by Panda et al. in their 2025 paper “ AccessEval: Benchmarking Disability Bias in Large Language Models” on arXiv [1]. Ableism bias in LLMs remains a small area of
The question I had been arguing about was grep versus vectors. The question that mattered was whether the store knows what kind of fact it is holding. A few weeks ago I published “Your Agent’s Memory Is Probably Wrong.” It opened with a filesystem agent using grep beating Mem0 on LoCoMo, and the whole piece leaned on that result. Then I wrote “Tail Rent,” about what happens to a long-running agent session after its third compaction, when a decision about a refresh token TTL had been sanded down into “auth token handling was reviewed and updated.” Both pieces were about memory going wrong. Both
OpenAI on Google Cloud A new model route is easy to request. Proving that it is safe, usable, affordable, and reversible is the real engineering work. A promising AI preview belongs behind evidence gates, not on an automatic path to production. One phrase turns an AI pilot into a future incident: “It worked in the preview.” It worked for one prompt, on one laptop, with an engineer watching. That says almost nothing about who can invoke it, what happens to sensitive context, whether the needed API behavior exists, what it will cost at volume, or how quickly the team can stop it. That distinctio
A skill that rewrites itself is a deployment. I want it to go through the same door as every other deployment. Note to self before publishing: every block marked [[ALI: ...]] needs real numbers from my own AutoHarness run. Do not publish with any of them still in place, and do not round them up. I have a folder of Claude Code skills in a .NET repository that I wrote by hand over the last few months. Some of them are good. Some of them are a pile of “and also, never do X” lines that I added the night something went wrong. A couple of them I am fairly sure nobody has triggered since July. So whe
One endpoint is convenient. One endpoint that quietly rewrites your model request is an outage waiting for a release you did not control. An agent works perfectly in a direct provider sandbox. It streams, calls tools, keeps a long context warm, and returns a strict object your application can trust. Then the team moves it behind an LLM gateway to centralize billing and policy. A new model setting disappears. A usage field no longer arrives. A streaming tool loop stalls. Nothing crashes loudly enough to make the cause obvious. This is the hidden trade-off in multi-provider AI: a gateway can mak
How I ported a Python eval-driven setup to C#, why my LLM judge gave a dangerous answer a perfect score, and what I changed in the rubric afterwards. The moment I decided to take evals on my .NET agent seriously was not dramatic. I changed one sentence in a system prompt, ran the agent by hand four times, liked what I saw, and shipped it. Two days later I wanted to know whether that change had made anything worse. The only way I had to find out was to run the agent again, against a model, and squint at the output. That is the expensive way to answer a cheap question. Most of what I wanted to k
AI Code Review Blast Radius A faster coding agent does not make every change equally safe. The teams that keep speed are learning to route review attention by risk, not by who -or what-wrote the code. Your AI coding agent opens a pull request that looks clean. Tests are green. The diff is only 180 lines. Then a reviewer notices that one of those lines changes how a tenant is selected before an authorization check. The patch was small; its blast radius was not. That is the review problem many engineering teams have now. Agents can create more changes than a small group of people can read with u
Your lakehouse doesn’t slow down because the data grew. It slows down because the file count did — and every one of those files is a tax your query planner pays before it reads a single byte. Part I | Part II | Part III | Part IV | Part V | Part VI | Part VII | Part VIII | Part IX | Part X | Part XI | Part XII | Part XIII | Part XIV | Part XV | Part XVI | Part XVII | Part XVIII The 58-Minute Query That Read Nothing Useful The Slack message landed at 9:41 PM, eleven minutes into the nightly reporting window: “finance dashboard is stuck again.” Not failed — stuck. The aggregate query over our ev
Inside EdgeRock-4B’s “Rock Memory Anchors” — and what they mean for anyone serving long-reasoning models Every engineer who has served a reasoning model knows the moment. The model is good, the answers are right, and then someone sends a 64K-token agent trace and the GPU's memory graph turns into a staircase. Reasoning models think out loud, and every token they think costs KV cache. On October 19, 2026, a team from Rock Matrix (岩芯矩阵, a unit of Chinese listed company 岩山科技), working with researchers at HKUST and CMU, open-sourced a model that takes a different route. EdgeRock-4B keeps its atten
A session graph, shortest paths and a knapsack kept every rule at half the tokens of a summary. Then the next task cost twice as much. I looked at the compactions in my own long Claude Code sessions. Each /compact read between 437k and 830k tokens, took 134 to 182 seconds, and produced a summary of 20k to 28k characters. That summary is everything the rest of the session knows about what came before. A model wrote it in one pass, deciding on its own what would matter later. Two things about that bothered me. First, a summary paraphrases. “Amounts are integer cents end to end” can come back as
Every Clock Says 10:10? How AI Actually Generates Images — Diffusion, Latent Space & Flow Matching (Every Number Measured) Why every AI clock snaps to 10:10, how a 32-channel VAE cuts attention 65,536×, Tweedie’s formula, Classifier-Free Guidance, and 4-step Rectified Flow measured on a 96 GB Blackwell GPU. By Latent AI Labs (@LatentAILabs) In Part 1 ( YouTube ), we saw how an LLM turns words into vectors and why it miscounts the "r"s in "strawberry". In Part 2 ( YouTube ), we measured how thinking models catch their own mistakes by writing a scratchpad. And in Part 3 ( YouTube ), we opened al
