12 slices · ~8m
Speculative Decoding & DeepSeek DSpark
How modern LLMs generate text up to 3x faster without dropping a single percentage of quality.
Distilled AI news and machine learning research. Bite-sized explanations to help you catch up or understand core concepts in minutes.
New twice a week
Hot Slice Just Dropped★Tokens Cheap · Insights Cheaper★Sliced Weekly From Zürich★2M Tokens Or Your Money Back★No GPT Wrappers · No Crust Wasted★
Hot Slice Just Dropped★Tokens Cheap · Insights Cheaper★Sliced Weekly From Zürich★2M Tokens Or Your Money Back★No GPT Wrappers · No Crust Wasted★
Today's slice
Speculative Decoding & DeepSeek DSpark
8m read·12 slices
How modern LLMs generate text up to 3x faster without dropping a single percentage of quality.
Read
It.
Tap anywhereIt.
Today's slice
Featured
Speculative Decoding & DeepSeek DSpark
8 min read·12 slices
How modern LLMs generate text up to 3x faster without dropping a single percentage of quality.
Read
It.
It.
Click anywhere
01 / fresh
Fresh From The Oven.
Weekly AI News.
12 slices · ~7m
DwarfStar: run frontier AI on your own machine
A hyper-focused, native inference engine built for one job — running DeepSeek V4 locally, even on hardware that technically can't fit it.
13 slices · ~10m
Figma best practices in the age of AI
AI agents can now read your Figma files directly via MCP. The canvas that worked for human designers breaks for machine consumption. Here is what to change.
12 slices · ~7m
Open knowledge format: Google's AI memory standard
A minimal YAML spec that turns scattered wikis and PDFs into a structured, Git-trackable knowledge base that LLMs can query without hallucinating the details.
10 slices · ~7m
Claude Fable 5: The hidden truth
Anthropic's most capable model ever has a secret playbook. The official system card reveals silent classifiers, activation steering, and a 30-day data retention loop you never see.
10 slices · ~6m
iii: The end of backend fragmentation
A WebSocket-based engine that replaces API gateways, message queues, cron daemons and AI agent scaffolding with one unified runtime.
10 slices · ~5m
Book-to-Skill: Compile any PDF into native Claude knowledge
An open-source compiler that turns any technical book into a structured Claude Code skill. Pay 4,000 tokens per session instead of 200,000.
9 slices · ~5m
Agent Memory That Actually Sticks
A fully local engine that cuts token usage by 61% and boosts agent task pass rates by 51%, without calling a single external API.
8 slices · ~5m
Markdown With Superpowers
One text file. Four output formats. No export pipelines, no tool switching, no compromises. Quarkdown turns plain Markdown into a typesetting language capable of math and logic.
9 slices · ~5m
Mirage: The Universal File System for AI Agents
How one open-source project gives AI agents a single root directory across every platform, database, and cloud service.
11 slices · ~8m
When AI Builds Itself: The Dawn of Recursive Self-Improvement
What happens when AI no longer relies on humans to get smarter? The answer is already happening in production labs.
That's the tray.
More @ 6 a.m.
02 / paper
Slice of Paper.
Academic Research, Simple.
6 slices · ~5m
SkillOpt: Self-Evolving Agent Skills
Microsoft's new system lets AI agents rewrite their own tools when the world changes.
7 slices · ~6m
Attention Is All You Need
The 2017 paper that killed sequential AI and sparked the LLM era.
7 slices · ~6m
BERT & GPT-1: The Fork
The 2018 split that divided AI into readers and writers.
6 slices · ~5m
Language Models are Few-Shot Learners
The 2020 paper that made prompting the new programming.
6 slices · ~5m
InstructGPT & Chinchilla: The Optimization
The two 2022 papers that proved raw scale was only half the answer.
6 slices · ~5m
The Sensory Shift: Native Multimodality
How Gemini 1.5 and GPT-4o taught AI to see, hear, and speak from a single brain.
6 slices · ~5m
The Reasoning Era
How DeepSeek-R1 and s1 moved the AI frontier from training compute to thinking time.
That's the tray.
More @ 6 a.m.
03 / homemade
How to Make AI at Home.
Zero to Hero Tutorials.
7 slices · ~5m
Mixture of experts: How AI models got smarter without getting slower
The secret architecture behind trillion-parameter models that stay fast and affordable.
10 slices · ~7m
Steering vectors: the DJ mixer inside every LLM
Prompting talks to the model. Steering vectors operate inside it.
8 slices · ~5m
Tokenization: How AI Reads Text
AI doesn't read words. It reads bricks of text called tokens.
8 slices · ~7m
Embeddings: How AI Understands Meaning
Words as coordinates. The trick behind every AI that feels smart.
4 slices · ~5m
Vector Databases: Semantic Search at Scale
From keyword matching to finding meaning. The database that understands context.
7 slices · ~6m
Neural Networks: The Engine of AI
Digital neurons, layered deep. The engine behind every AI breakthrough.
5 slices · ~4m
Attention Mechanism & Transformers
The architectural secret behind GPT, Claude, and Gemini.
6 slices · ~5m
Weights & Biases: AI's Control Panel
70 billion dials. All tuned to perfection.
7 slices · ~5m
Pre-Training vs. Fine-Tuning
Two phases. One brain. Here's how a model goes from clueless to expert.
6 slices · ~5m
AI Alignment & RLHF
Why brilliant AI can still go terribly wrong. And how we fix it.
10 slices · ~8m
MCP: The Universal Adapter for AI
One open standard. Every data source. No more custom integration code.
That's the tray.
More @ 6 a.m.
04 / sauce
Secret Sauce.
Prompt Engineering Hacks.
12 slices · ~9m
The Art of Loop Engineering
How stacking loops turns unpredictable LLMs into production-grade systems.
15 slices · ~9m
The LangChain stack
From raw model to production AI. Here's the full toolkit.
11 slices · ~8m
Heretic: the abliterated AI models
What happens when AI loses its corporate moral compass? Inside the uncensored mechanics of abliterated neural networks.
7 slices · ~5m
Caveman
Fewer words. Fewer tokens. Bigger savings.
7 slices · ~7m
Spec-Driven Context Engineering
Stop context rot. Ship with the GSD Method.
8 slices · ~7m
The Autonomous Hill-Climber
What if the AI stubbornly fixed its own mistakes until it worked?
7 slices · ~6m
Prompt Caching
Ask a million times. Pay for one.
8 slices · ~5m
Progressive Context Disclosure
Give the AI the rules it needs. Nothing more.
That's the tray.
More @ 6 a.m.
Meet The Pizzaiolo
05 / About
Ciao, I'm The Cook.
SWE based in Zürich. Currently building next-generation products at Swisscom, including blue TV and myAI . Passionate about late-night coding and diving into research papers at 3 a.m.
- 7+ years of moving bits
- Lazy SWE
- Red Bull + coding
- Pizza lover (no pineapple)
06 / House Rules
How We Cook.
AI tokens get cheaper every quarter. Human attention doesn't. We prioritize the second: cheap context, expensive thinking. Every slice is crafted for high density: one clear concept at a time, with zero fluff.