Impact Of On-Chip SRAM Size And Frequency On Energy Efficiency And Performance of LLM Inference (Uppsala Univ.)

A new technical paper titled “Prefill vs. Decode Bottlenecks: SRAM-Frequency Tradeoffs and the Memory-Bandwidth Ceiling” was published by researchers at Uppsala University. Abstract “Energy consumption dictates the cost and environmental impact of deploying Large Language…

Our simulation methodology combines OpenRAM for energy modeling, LLMCompass for latency simulation, and ScaleSIM for systolic array operational intensity.

Counter-intuitively, high compute frequency can reduce total energy by reducing execution time and consequently decreasing static energy consumption more than the resulting dynamic power increase

The Big LLM Architecture Comparison

From DeepSeek-V3 to Kimi K2: A Look At Modern LLM Architecture Design

Ollama Search

Search for models on Ollama.

Nano Banana Pro is the best AI image generator, with caveats

The problem with Nano Banana Pro is that it’s too good.

Nano Banana can be prompt engineered for extremely nuanced AI image generation

Nano Banana allows 32,768 input tokens and I’m going to try to use them all dammit.

Proximal Policy Optimization

We’re releasing a new class of reinforcement learning algorithms, Proximal Policy Optimization (PPO), which perform comparably or better than state-of-the-art approaches while being much simpler to implement and tune. PPO has become the default reinforcement learning algorithm…

LLM Research Papers: The 2025 List (July to December)

In June, I shared a bonus article with my curated and bookmarked research paper lists to the paid subscribers who make this Substack possible.

2025: The year in LLMs

This is the third in my annual series reviewing everything that happened in the LLM space over the past 12 months. For previous years see Stuff we figured out about …

It turned out that the real unlock of reasoning was in driving tools.

notable result is that AI assisted search actually works now.

LLM Research Papers: The 2025 List (July to December)

A curated list of LLM research papers from July–December 2025, organized by reasoning models, inference-time scaling, architectures, training efficiency, and...

Reasoning

Reinforcement Learning

The State of Reinforcement Learning for LLM Reasoning

Understanding GRPO and New Insights from Reasoning Model Papers

the muted response to GPT-4.5 and Llama 4 (non-reasoning) models suggests we are approaching the limits of what scaling model size and data alone can achieve.

The State Of LLMs 2025: Progress, Problems, and Predictions

A 2025 review of large language models, from DeepSeek R1 and RLVR to inference-time scaling, benchmarks, architectures, and predictions for 2026.

DeepSeek released their R1 paper in January 2025, which showed that reasoning-like behavior can be developed with reinforcement learning, it was a really big deal

DeepSeek R1 was released as an open-weight model

Documentation

Claude API Documentation

Building an AI agent inside a 7-year old Rails application

We run a multi-tenant Rails application with sensitive data and layered authorization. In this post, I walk through how I added the first AI agent tool using RubyLLM, Pundit policies, and our existing Algolia search, without introducing a parallel system or loosening constraints.

LangGraph Explained from Scratch | Aman Kharwal

In this article, I’ll walk you through a complete guide to LangGraph from the ground up. LangGraph Explained from Scratch.

In LangChain, this usually means moving from user input to a prompt, then to the LLM, and finally to the output

ollama pull llama3

2025 LLM Year in Review

2025 Year in Review of LLM paradigm changes

Reinforce

Reinforcement Learning from Verifiable Rewards (RLVR)

Prompt caching: 10x cheaper LLM tokens, but how? | ngrok blog

A far more detailed explanation of prompt caching than anyone asked for.

The Behavioral Economics Of AI Agents - Why Consistency Doesn't Equal Trust - Dataconomy

The AI industry has a trust problem that mirrors a paradox Daniel Kahneman identified decades ago in human decision-making: people

Mistral 3 Live!

Frontier AI by hand ✍️

The 70% factuality ceiling: why Google’s new ‘FACTS’ benchmark is a wake-up call

rbenv and CLI executables

Meet Gemini, Google’s AI assistant. Get help with writing, planning, brainstorming, and more. Experience the power of generative AI.

The Concise Guide to Perplexity

Practical Guide on how to build an Agent from scratch with Gemini 3

A step-by-step practical guide on building AI agents using Gemini 3 Pro, covering tool integration, context management, and best practices for creating effective and reliable agents.

The DoorDash problem and the great AI browser fight

Amazon sued Perplexity this month over its Comet browser, which uses AI agents to do online shopping on your behalf. This is the first major front in the war over who gets to browse the web.

Comparing Memory Systems for LLM Agents: Vector, Graph, and Event Logs

Learn how different memory systems affect multi-agent planning. Comparing Memory Systems for LLM Agents highlights key performance metrics.

Plain Vector RAG

sublinear scaling

Comparing the Top 7 Large Language Models LLMs/Systems for Coding in 2025

Compare the top 7 large language models and systems for coding in 2025. Discover which ones excel for software engineering tasks.

OpenAI GPT-5 / GPT-5-Codex Anthropic Claude 3.5 Sonnet / Claude 4.x Sonnet with Claude Code Google Gemini 2.5 Pro Meta Llama 3.1 405B Instruct DeepSeek-V2.5-1210 (with DeepSeek-V3 as the successor) Alibaba Qwen2.5-Coder-32B-Instruct Mistral Codestral 25.01

Core coding quality: HumanEval, MBPP / MBPP EvalPlus, code generation and repair quality on standard Python tasks. Repo and bug-fix performance: SWE-bench Verified (real GitHub issues), Aider Polyglot (whole-file edits), RepoBench, LiveCodeBench. Context and long-context behavior: Documented…

In a First, AI Models Analyze Language As Well As a Human Expert | Quanta Magazine

If language is what makes us human, what does it mean now that large language models have gained “metalinguistic” abilities?

How I Use Every Claude Code Feature

A brain dump of all the ways I've been using Claude Code.

commish app design

A conversational AI system that listens, learns, and challenges

The Big LLM Architecture Comparison

From DeepSeek-V3 to Kimi K2: A Look At Modern LLM Architecture Design

this section, I’ll focus on two key architectural techniques introduced in DeepSeek V3 that improved its computational efficiency and distinguish it from many other LLMs:

Multi-Head Latent Attention (MLA)

Marktechpost/AI-Tutorial-Codes-Included: Codes/Notebooks for AI Projects

Codes/Notebooks for AI Projects. Contribute to Marktechpost/AI-Tutorial-Codes-Included development by creating an account on GitHub.

5 Common LLM Parameters Explained with Examples

Learn the 5 common LLM parameters explained with examples to optimize your model's performance and generate desired results.

max_completion_tokens, temperature, top_p, presence_penalty, and frequency_penalty

Max Tokens is the maximum number of tokens the model can generate during a run

ChatGPT's Atlas: The Browser That's Anti-Web

A blog about making culture. Since 1999.

Living dangerously with Claude

I gave a talk last night at Claude Code Anonymous in San Francisco, the unofficial meetup for coding agent enthusiasts. I decided to talk about a dichotomy I’ve been struggling …

Qwen3-VL· Ollama Blog

Ollama now supports Alibaba's Qwen3-VL.

https://venturebeat.com/ai/researchers-find-adding-this-one-simple-sentence-to-prompts-makes-ai-models

Verbalized Sampling (VS)

Generate 5 responses with their corresponding probabilities, sampled from the full distribution

Just Talk To It - the no-bs Way of Agentic Engineering | Peter Steinberger

A practical guide to working with AI coding agents without the hype.

Machine Learning Mastery

Making developers awesome at machine learning.

A Guide to Fine-Tuning LLMs using LoRA

In this article, I'll take you through a step-by-step guide to fine-tuning LLMs with LoRA. A Guide to Fine-Tuning LLMs using LoRA.

7 LLM Generation Parameters—What They Do and How to Tune Them?

Seven LLM generation parameters: max tokens, temperature, top-p, top-k, penalties, stop sequences, tuning guidance, defaults

stop sequences

presence penalties

OpenAI’s Windows Play

OpenAI is making a play to be the Windows of AI: the all-encompassing platform that controls both hardware supplier and software developers.

Embracing the parallel coding agent lifestyle

Plus prompt injection attacks against Sora 2 cameos and notes on DSPy and Litestream 0.5.0

Build a Reasoning Model (From Scratch) - Sebastian Raschka

Understand LLM reasoning by creating your own reasoning model–from scratch! LLM reasoning models have the power to tackle truly challenging problems that require finding the right path through multiple steps. In Build A Reasoning Model (From Scratch) you’ll learn how to build a…

Understanding the 4 Main Approaches to LLM Evaluation (From Scratch)

Multiple-Choice Benchmarks, Verifiers, Leaderboards, and LLM Judges with Code Examples

four common ways of evaluating trained LLMs in practice: multiple choice, verifiers, leaderboards, and LLM judges

training loss, perplexity, and rewards, but they are usually used internally during model development.)

A Coding Guide to Build an Autonomous Agentic AI for Time Series Forecasting with Darts and Hugging Face

A Coding Guide to Build an Autonomous Agentic AI for Time Series Forecasting with Darts and Hugging Face. A Step by step guide

LoRA Without Regret

How LoRA matches full training performance more broadly than expected

OpenAI’s Sora Makes Disinformation Extremely Easy and Extremely Real

The new A.I. app generated videos of store robberies and home intrusions — even bomb explosions on city streets — that never happened.

Building the heap: racking 30 petabytes of hard drives for pretraining

How we spent under half a million dollars to build a 30 petabyte data storage cluster in downtown San Francisco

Top 10 Local LLMs (2025): Context Windows, VRAM Targets, and Licenses Compared

Compare top local LLMs for 2025: context windows, VRAM tiers, licenses, quantization, runtimes, dense vs MoE tradeoffs clearly

5 Cutting-Edge Natural Language Processing Trends Shaping 2026

In this article, we discuss five cutting-edge NLP trends that will shape 2026.

transformers’ biggest weakness remains the high compute time and memory consumption of self-attention

Approaches like linear attention and sparse attention

GPT-5-Codex

OpenAI half-relased this model earlier this month, adding it to their Codex CLI tool but not their API. Today they've fixed that - the new model can now be accessed …