Step-By-Step LLM Engineering Projects Roadmap – Build a tokenizer
– Learn embeddings
– Implement RoPE / ALiBi
– Hand-wire attention
– Build MHA
– Build a Transformer block
– Train a mini-former
– Compare objectives
– Build sampling
– Speculative decoding
– KV cache
– MQA / GQA /
CODE
-
Step-by-step LLM engineering projects roadmap
By
–
-
Scaling Past Informal AI: Math, Lean, and formal proofs for AGI
By
–
Scaling Past Informal AI https://
latent.space/p/axiom @axiommathai founder & CEO @CarinaLHong explains why math may be the missing path from code agents to AGI, why verified AI is about scaling brilliance not just fixing hallucinations, how Lean and formal proofs turn reasoning -
Compute usage: 1B ChatGPT vs 5M Codex users
By
–
Who is using more compute – 1b of ChatGPT users or 5m of Codex users?
-
Codex tip: use plugins to boost power, just ask Codex
By
–
Codex tip: one of the easiest ways to make Codex much more powerful is openai/plugins
there are now 100+ plugins for everything from building apps to working with Figma, Notion, Vercel, and more the best part: you do not even need to know which ones exist
open Codex and ask: -
OpenAI adds new capabilities to GPT-Rosalind for drug discovery
By
–
We’re bringing new capabilities to GPT-Rosalind, a model series purpose-built for life sciences research at enterprise scale. It brings GPT-5.5’s agentic coding and tool use together with stronger intelligence for drug discovery, analysis, design, and experimental workflows.
-
Claude Code spawns hundreds of AI agents from one prompt
By
–
Claude Code can now spawn hundreds of AI agents from a single prompt. It writes its own orchestration script, breaks your task into subtasks, and runs them all in parallel. One developer used it to port 750,000 lines of code in 11 days. Here's how dynamic workflows actually
-
LFM2/2.5 architecture: convolution is almost all you need
By
–
Oh fun! The LFM2/2.5 architecture is a joy to play with. Convolution is (almost) all you need.
-
Benchtalks #2 discusses ProgramBench where frontier models scored 0%
By
–
Benchtalks #2 is up with @vincentsunnchen. @jyangballin of @stanfordnlp, creator of @SWEbench, on ProgramBench, the benchmark every frontier model scored 0% on at launch.
— Snorkel AI (@SnorkelAI) 3 juin 2026
They dive into end-to-end code generation, why models reward-hack once they get internet access, and the… https://t.co/WZmhUqa8yaBenchtalks #2 is up with @vincentsunnchen
. @jyangballin of @stanfordnlp
, creator of @SWEbench
, on ProgramBench, the benchmark every frontier model scored 0% on at launch. They dive into end-to-end code generation, why models reward-hack once they get internet access, and the -
AI-Enabled Cyberattacks vs. Security Community Techniques: An Analysis
By
–
How well do the security community's techniques hold up against AI-enabled cyberattacks? We examined 832 malicious accounts and mapped their activity onto a longstanding database of tactics and techniques used by threat actors. Here's what we learned:
-
Gemma 4 hits 150M+ downloads with new, powerful 12B model for local use
By
–
Celebrating the milestone of a massive 150+ million downloads of Gemma 4 with the release of the new Gemma 4 12B model! It's incredibly powerful for such a small model and it’s tiny enough to run locally on a laptop with just 16GB VRAM. Apache 2.0 license – happy building!