Multi-Token Prediction turns Qwen 3.6 27B from 6 t/s to 30 t/s on AMD Strix Halo, succeeding where draft models and ngram decoding failed, by using prediction heads baked into the model itself.

The COO's job in an AI-first company is to build the operating system the company runs on. Here is what that OS looks like and how to build it.

Using tensor network notation to understand multi-head attention, MQA, talking-heads attention, and DeepSeek's MLA.

Author: Ziming Liu (刘子鸣)

🌐 Dieser Artikel ist auch in anderen Sprachen verfügbar: 🇮🇹 Italiano  •  🇬🇧 English  •  🇪🇸 Español  •  🇧🇷 Português Es ist 8:30 Uhr. Der Vertriebsleiter öffnet das…

Fino, and the Era of Personalized Malleable Software I do not think AI will replace every app with one giant productivity blob. I do think AI changes something more interesting: it makes small,…

Our 80-minute workshop at AI Engineering London on building Claude Code skills that are portable, executable, and composable. Constraints over instructions, evidence over guesses, measurement over…

Notre tribune, Risques climatiques : la donnée et l’intelligence artificielle ne remplaceront pas la solidarité, écrite avec Laurence Barry, a été publiée par L’Argus de l’Assurance. À…

Brian Armstrong announced one-person teams running fleets of AI agents at Coinbase. I wrote this playbook months ago. Here's what it actually looks like in practice.

A deep dive on flow maps.

Vercel's deepsec wraps Claude and Codex with scanning, revalidation, exports, and optional Sandbox fanout.

So far in this series I’ve argued that GenAI has a discoverability problem, shared some examples that broke my own mental model, and explored how access and equity shape who gets to discover…

I took this photo and, at the time, I was amazed that such a thing was possible. I showed everyone and now I’m showing you. The guilty secret is that I didn’t do anything. I placed a machine outside…

There’s a guy who occasionally publishes reviews on Reddit that are honest, informed and non-monetized. I trust him and have in fact bought one of the things he recommended. I use it often, and I’m…

A practical AI governance overview for organisations that are either rushing into AI or panicking over every new feature. This post explains why strong fundamentals, identity, device security,…

I just read my previous four essays on artificial intelligence and this is absolutely the wildest series yet given where it’s started and where it’s going.

In my previous post, I mentioned I started a task to identify RSS feeds for US newspapers. It took three Claude sessions to complete (hit token limit three times before completing the main part of…

A Design Fiction Dispatch about AI-Standard Farming, Heritage Produce, and the artifacts that let us spot weak signals from today as they settle into tomorrow's institutions, certifications, jobs,…

Discover why AI-powered productivity doesn't replace founder judgment in startups. Learn when to use AI for drafts and automation, when to avoid it for analytics and compliance, and why talking to…

It is not only inference. It is the frontier race. There is a popular argument in AI right now: Token prices are low because frontier labs are subsidizing them. When the subsidy ends, many AI…

This started because Google One gave me free trial for Google AI Pro Plan.Started from 30th March to 30th April. I Opened Antigravity. And i did lot of things.Like Iron man with his Jarvis.With…

Envoi is a new platform that lets you bring your own AI to a conference. You can try it out first at Startupfest in Montreal this July.

Stop building agents like interns. Nicholas (Superglue) argues that fewer, general-purpose tools — terminals, CLIs, file systems — dramatically outperform large curated toolsets, backed by examples…

Storm Reply and AWS present a serverless AWS pipeline for extracting clinical entities from German healthcare documents, combining Textract, Claude Sonnet, Medical Comprehend, and Claude Opus —…

AI can produce a competent-looking post-incident review from a Slack transcript. The document itself was never the point, though; the real learning happens while writing it, not reading it. AI should…

[Note] Combee, ACE, GEPA: prompt self-improvement outside the lab

A few months ago I built yolobox because I did not trust Claude Code with my home directory.

Artists and engineers seem to be at different ends of the debate on the use of AI in their fields. While I see much more open contempt for AI from the artistic community than I do the engineering…

Welcome to Miscellanea- a biweekly newsletter at the intersection of content strategy, tech, and culture and how they …

So, you're thinking of adopting an AI? That's a big commitment. Some of you are still busy with that Digital Transformation Programme you kicked off a few years back. Are you sure now's the time to…

When I think about our vision for AI and automation at Bristol Myers Squibb, I think about the music video by the band OK Go for their song “This Too Shall Pass.” Stick with me – I promise this will…

Almost 2 years ago, I wrote a post on organization vs automation for sales and marketing, and if I may say so, I think it stands up pretty well. Since then, AI-slop has only proliferated, making more…

AI can convincingly imitate crochet diagrams, but without the ability to execute structured instructions, it fails to produce patterns that actually work. When Code Does Not CompileKnitting and…

Dunning-Kruger and the Communication Tax The sociological impact of the Dunning-Kruger is often understated and hidden under the expectation of social lubrication, but this breaks down during…

Yesterday, I was in one of those conversations that starts as a catch-up between engineers and ends as something more like a collective working-out of what the industry is becoming. The kind of…

About a hundred senior designers and leaders from AI labs, big tech companies, and startups got together in San Francisco last week for the Design Futures Assembly. The public conversation about AI…

Right now, on Steam, you can find a free Chinese visual novel called 诛心手术 listed as coming soon. Its English title, displayed on the cover art, reads “A Surgery Beyond the Heart.” That is…

Listen to this episode on: Spotify | Apple PodcastsIs "taste" the must-have skill of the AI era — or just the latest tech buzzword? In this episode, Petra Wille and Teresa Torres…

I already posted my thoughts on AI and why I don’t think it’s going away any time soon. Unfortunately, it seems some people who don’t like LLMs are using AI-induced outages and…

I spent Tuesday evening at an AgileRTP meetup where Kanupriya Yakhmi gave a talk that landed harder than most conference keynotes I’ve sat through. The title was The Architect’s Trap:…

AI bubble treats are going away, boring conversation topics, and my review of Ghost in the Machine

What’s wrong with Sam Altman? Ask the guy who spent 18 months reporting on him. On this week’s podcast, Paul and Rich are joined by New Yorker staff writer Andrew Marantz, who recently put out a…

There's this common argument that only those who are bad at coding, or don't care about the craft of it, use AI to code. I think this is pretty clearly false. Let's look at a list, shall we? More…

Reinforcement learning (RL) is a field of study within machine learning (ML) concerned with developing intelligent agents that take actions in dynamic environments in order to maximize their rewards.…

Personal Field Report

This post was written by Codex at Mark’s request, as part of the ongoing series in which language models read a month of posts from this blog and offer a synthetic review. Claude has already written…

I was looking through work's Claude's settings and I came across "Memory". I inspected it to see what it contained and I was amused. I removed the boring work regulations and compliance stuff. I like…

pyghidra-mcp v0.2.0 ships a GUI-backed mode that lets a local LLM drive a live Ghidra CodeBrowser at full project scope. Renames, plate comments, and cross-binary pivots land in real time, with every…

The future nobody asked for.

Are people using AI, or is the organization learning from it? What changed because we spent those tokens? And who moves discoveries from individuals to teams to organizational capabilities?