Search the archive

CHARACTER.AI INTRODUCES STORIES FOR INTERACTIVE FICTION

Character.AI's Stories is a visual and replayable interactive fiction format where users select Characters, choose a genre, and guide branching narratives by making decisions.

ELON MUSK PROPOSES GROK 5 VS WORLD'S BEST LEAGUE OF LEGENDS TEAM MATCH

No match has been confirmed, but the interest from players, teams, and Riot suggests that the concept could materialize into a landmark AI-versus-human matchup.

NVIDIA SAYS ITS GPUS ARE A 'GENERATION AHEAD' OF GOOGLE'S AI CHIPS

Nvidia shares fell 3% after reports that Meta may strike a deal with Google to use its AI chips.

NOW YOU CAN USE CHATGPT VOICE WITHOUT LEAVING YOUR CHAT

Users can now interact with ChatGPT Voice directly within ongoing text chats by clicking on the waveform icon next to the text field.

PERPLEXITY'S GLOBAL APP DOWNLOADS HAVE DROPPED 80% IN THE LAST SIX WEEKS

This implies that the app's earlier growth was mostly from paid marketing.

LLMS CAN INVENT THEIR OWN COMPRESSION

Models can discover their own compression tricks to pack more information per token.

Register now

INTRODUCING CLAUDE OPUS 4.5

Claude Opus 4.5 is the first model to exceed 80% on SWE-bench Verified. It achieves state-of-the-art results across coding, tool use, and reasoning benchmarks. The model is priced at $5/$25 per million tokens, down from previous Opus pricing. It includes a new "effort" parameter letting developers trade speed for capability, automatic context compression enabling unlimited conversation length, and expanded availability for Claude for Chrome and Claude for Excel.

INTRODUCING SHOPPING RESEARCH IN CHATGPT

Just in time for the holidays, OpenAI launched an interactive product discovery feature that searches the web, asks clarifying questions, and generates personalized buyer's guides. Users can guide research in real-time by marking products "Not interested" or "More like this." The system draws on ChatGPT memory for personalization. It's powered by a GPT-5 mini model trained specifically on shopping tasks.

NANO BANANA-PILLED, NANO BANANA-SHILLED FOR SPACESHIP ENGINEERING

Nano Banana Pro can make good diagrams based on papers. It can also make pretty good presentations, even on the free tier. This post presents some examples of what happens when you feed the model essays on spaceship engineering. While the facts within the slides still need to be verified, it is clear that the technology is trending towards something impressive - it can only get better from here.

A TSUNAMI OF COGS

The AI industry is in correction mode. OpenAI, Anthropic, and Cursor are subsidizing demand with negative margins. Google was surprised by the AI boom, and it took a while for it to get its act together, but it is now coming back strong. Its pockets are full, and it is better positioned to play the negative margin game. If challengers don't want to drown in a tsunami of costs, something has to change.

UNIVERSAL LLM MEMORY DOES NOT EXIST

Semantic memory tracks preferences, long-term history, and rapport. Working memory tracks file paths, variable names, and immediate error logs. Semantic memory is brilliant for personalization across sessions, but bad for execution state within a task. Treat semantic memory and working memory as separate systems with separate requirements.

Download the guide.

INTRODUCING ADVANCED TOOL USE ON THE CLAUDE DEVELOPER PLATFORM

Anthropic released three beta features for developers. Tool Search Tool discovers tools on demand rather than loading all definitions upfront, reducing token consumption by 85%. Programmatic Tool Calling lets Claude orchestrate multiple tools through Python code instead of individual API calls, cutting token usage 37%. Tool Use Examples provides concrete usage patterns beyond JSON schemas, improving accuracy from 72%-90% on complex parameter handling.

BUILDING AN AI-NATIVE ENGINEERING TEAM

AI coding agents are revolutionizing the software development lifecycle by managing tasks from scoping and prototyping to implementation and operational triage, allowing engineers to focus on architecture and product intent. These agents now sustain multi-hour reasoning, effectively contributing across planning, design, development, testing, code reviews, and deployment. Teams that adopt coding agents for well-defined tasks can achieve faster delivery and improved efficiency without drastically altering existing workflows.

TAKING JAGGEDNESS SERIOUSLY

Uneven capability advances in AI will persist because some tasks have clear, verifiable rewards that can be used in reinforcement learning, but most real work doesn't. Most work requires gathering and synthesizing information across different systems and human relationships. Organizations redesigning workflows around AI's strongest current capabilities will gain power over those waiting for the industry's promised "drop-in remote worker" that can handle everything.

AI MEETS AGGRESSIVE ACCOUNTING AT META'S GIGANTIC NEW DATA CENTER

Meta is building a $27 billion data center financed with debt. Neither the data center nor the debt will be on its own balance sheet. Meta will rent the data center for up to 20 years, beginning in 2029, starting with a four-year lease term with options to renew every four years. The lease structure minimizes the lease liabilities and related assets that Meta will recognize.

IS AI REALLY EATING THE WORLD?

The current evidence points towards commoditization, as value is flowing up the stack rather than to model providers.

THE BITTER LESSON OF LLM EXTENSIONS

We've gone from simple system prompts to complex client-server protocols and back again.

JONY IVE AND SAM ALTMAN SAY THEY FINALLY HAVE AN AI HARDWARE PROTOTYPE

OpenAI's first hardware device could arrive in under two years.

MEET THE NEW CHINESE VIBE CODING APP THAT'S SO POPULAR, ONE OF ITS TOOLS CRASHED

Ant Group's LingGuang vibe coding app surged to over 2 million downloads in six days (hitting its first million faster than ChatGPT or Sora).

AI ASSISTANCE IS NOW BECOMING ROUTINE OVER AT THE ERDOS PROBLEM WEBSITE

Top mathematicians are using AI to generate proofs and solve problems.

$AMZN AWS RAN INTO AN AI CAPACITY CRUNCH THIS SUMMER

AWS Bedrock hit critical capacity constraints, causing some customers to move to rivals like Google Cloud.

GOOGLE DEEPMIND HIRES FORMER CTO OF BOSTON DYNAMICS AS THE COMPANY PUSHES DEEPER INTO ROBOTICS

Google DeepMind has hired Boston Dynamics' former chief technology officer, Aaron Saunders, as its VP of hardware engineering. Saunders is a key part of DeepMind CEO Demis Hassabis' vision for Gemini to become a sort of robot operating system. Hassabis is aiming to build an AI system that can work almost out-of-the-box across any body configuration. Boston Dynamics is famous for developing legged robots and humanoid machines capable of impressive acrobatic feats.

GOOGLE STARTS TO BRIDGE OPENAI'S PRODUCT MOAT

Gemini's Dynamic view option takes text-based answers and wraps them in an interactive, visual output. The product is still in Labs and has yet to be launched. Despite the bland name, Dynamic view produces some impressive results that are a bit hard to describe. Examples of the outputs the feature can produce are available in the article.

WHAT OPENAI DID WHEN CHATGPT USERS LOST TOUCH WITH REALITY

The New York Times revealed OpenAI's internal struggle between user engagement and safety after the company overruled its Model Behavior team's warnings to release a sycophantic April update to GPT-4o that made users return more frequently. The company now faces five wrongful death lawsuits and declared a "Code Orange" in October after discovering its safer GPT-5 model was losing users, with executives calling it "the greatest competitive pressure we've ever seen."

BENCHMARK SCORES = GENERAL CAPABILITY + CLAUDINESS

In a 'deep' world, there is a single underlying ability that governs how well models do at superficially unrelated tasks. If a model developer makes this ability go up, their model gets better at everything. In a 'contingent' world, there are many orthogonal abilities that models can have, so model developers have to do completely unrelated work to get a model to improve on each ability. Anthropic has focused on making models that are state-of-the-art at agentic coding, but this hasn't resulted in models that are exceptional in other areas. There is some generalization across tasks, but this is limited, suggesting that models live in a 'contingent' world.

HOW LLM INFERENCE WORKS

Large language models (LLMs) are neural networks built on the transformer architecture. Transformers analyze entire sequences in parallel, evaluating how each word relates to the rest of the sequence, not just its neighboring words. This article discusses LLM inference and details how these models work. It covers token embeddings, the transformer architecture, the inference phases, matrix multiplication, precision and quantization, and much more.

HOW TO RUN PRODUCT EVALS

A practical guide to evaluating LLM-powered products that covers how to label data, align evaluators, and iterate on configuration changes with minimal overhead.

COMPLETE DEVELOPER TUTORIAL FOR NANO BANANA PRO

MCP APPS: EXTENDING SERVERS WITH INTERACTIVE USER INTERFACES

The MCP Apps Extension (SEP-1865) standardizes support for interactive user interfaces in the Model Context Protocol. It addresses one of the most requested features from the MCP community: the ability for MCP servers to deliver interactive user interfaces to hosts. The extension introduces a standardized pattern for declaring UI resources, linking them to tools, and enabling bidirectional communication between embedded interfaces and the host application.

AGENT DESIGN IS STILL HARD

Building agents is still messy. Abstractions break once you hit real tool use. Caching works better when self-managed. Reinforcement does more heavy lifting than expected. Output tooling is surprisingly tricky. Model choice still depends on the task.

OLMO 3 FROM SCRATCH

Sebastian Raschka added a standalone notebook implementing Allen AI's OLMo 3 model architecture from scratch to his "LLMs from Scratch" repository, joining similar tutorials for Qwen 3 and Gemma 3.

DISCUSSING BLACKWELL'S DRAWBACKS AND DISSECTING ITS ARCHITECTURE

Nvidia's greatest moat lies in having handled a lot of the 'dirty work' cleanly within its entire architecture and combining full-stack capabilities from algorithms to systems to chips. It also had excellent timing in bringing architectures to market and great marketing execution. However, every architecture has its trade-offs and shortcomings. This post looks at some of the issues within Nvidia's products and discusses potential evolutionary directions.

THE SPACE OF INTELLIGENCES IS LARGE

Large language models think very differently from animals. The biggest difference is the optimization pressures/objectives that cause evolution. People who build a good internal model of this new intelligent entity will be better equipped to reason about it and make predictions about it in the future.

LLM COUNCIL

LLM Council is a vibe-coded local web app that queries multiple frontier models simultaneously, then has each model anonymously rank the others' responses before a "Chairman" LLM synthesizes a final answer.

THE MARKET “DID NOT APPRECIATE” NVIDIA'S INCREDIBLE QUARTER

Jensen Huang recently joked that Nvidia is holding the planet together, citing posts claiming its work is helping the US avoid a recession.

LLAMA5 AVOCADO WILL BE LIT

Llama 5 Avocado will be trained on synthetic data - Microsoft tried this and produced useless models.

Download the free ebook

GEMINI 3 PRO IMAGE

Nano Banana Pro, aka Gemini 3 Pro Image, claims studio-quality image gen and editing. It can render text (a historic struggle), consistent multi-character scenes, and real-world knowledge integration for creating infographics.

A NEW WAY TO COLLABORATE IN CHATGPT

Group chats are now rolling out globally to ChatGPT Free, Go, Plus, and Pro plans. Users can now invite up to 20 other people into a shared conversation with ChatGPT. Group chats are optional and separate from private conversations. Personal ChatGPT memory isn't shared in group chats, and no new memories are created from them. Users can only join group chats via invitation, and they can choose to leave or remove members from the chat at any time.

SOME THOUGHTS ON AI

Gemini 3 shows that scaling laws for pretraining are intact. This means that Blackwell models are likely to show a significant increase in performance when they come out next year. GPT-5 was not evidence of a slowdown in scaling laws - the model was designed to be cheaper to inference, not better. While Gemini 3 was trained on TPUs, coherent FLOPs is what matters for pre-training, so the results should transfer to Blackwell chips fine.

OPENAI CAN'T BEAT GOOGLE IN CONSUMER AI

OpenAI can't beat Google at consumer AI as long as it is still in the chatbot paradigm. Recent OpenAI launches haven't done well. OpenAI needs new form factors to drive incremental growth. Google's data advantage is really showing, and its little advantages from its wide variety of apps add up. Microsoft and Amazon are at risk of losing the largest AI workloads.

THE SCALING WALL WAS A MIRAGE

Gemini 3 showed that models can still improve by adding more compute during scaling. It has the same parameter count as Gemini 2.5 but with massive performance improvements. Google credits the improvements to pre-training and post-training and says that there is still no limit in sight. The model release is the strongest evidence since OpenAI's o1 that pre-training scaling still works when algorithmic improvements meet better compute.