The Tokenpocalypse Is Here: Companies Are Scrambling To Stop Spending So Much on AI

Simon Willison · 2026-08-07 · 2 min read

Companies are burning through LLM tokens at unexpected scale not because of engineering workloads but because non-engineers are doing things like converting PDFs to images then to markdown before feeding them to AI models, a needlessly token-heavy pipeline. For a senior engineer, this is a concrete reminder that AI cost optimization requires governance and guardrails around how data is preprocessed and passed to models, not just prompt efficiency.

What is this #Agent #Plugins 1.0 spec #ai #agenticengineering #codex

Agentic Engineering · 2026-08-07

This video introduces the new Agent Plugins 1.0 specification, an open standard designed to package reusable instructions and tool connections for AI agents. It aims to standardize how agent capabilities are bundled and moved across different AI agent clients (e.g., Copilot to Codex) without requiring repetitive re-packaging or re-configuration. The specification focuses on creating a small, common package boundary—a directory with a `plugin.json` and a skills folder—rather than standardizing the agents themselves or defining installation, distribution, or security models.

DeepMind Just Changed How AI Sees The World

Two Minute Papers · 2026-08-07

Google DeepMind has introduced Gemma 4, a family of open-weight, natively multimodal language models that redefines efficiency and capability. Unlike traditional AI systems that use separate neural networks for different modalities (e.g., vision, audio), Gemma 4 integrates raw sensory input directly into its main transformer, blurring the lines between perception and thinking. This innovative architecture results in significantly smaller models (up to 99% smaller) that can run locally on devices like laptops, yet outperform much larger, specialized models in complex tasks involving multiple modalities.

read more

Google DeepMind's Gemma 4 represents a significant leap in multimodal AI, primarily by introducing a unified architecture that handles various sensory inputs (like images and audio) directly within a single main transformer. This contrasts sharply with conventional approaches where a large language model (LLM) acts as a central hub, but each modality (vision, audio) requires its own dedicated encoder neural network to preprocess raw input into a format the LLM can understand. For example, in a conventional setup, an image would first pass through a vision model (e.g., a Vision Transformer) to be translated into meaningful tokens, and similarly, an audio clip would be processed by an audio encoder.

The core innovation of Gemma 4 is its ability to bypass these specialized encoders and feed raw sensory data directly into its main transformer. For visual input, an image is divided into small patches (e.g., 16x16 pixels). Instead of using a separate Vision Transformer to convert these patches into embeddings, Gemma 4 directly projects these pixel patches into its internal representation, preserving their positional information. Similarly, for audio, the raw waveform is sliced into 40-millisecond chunks, and these raw chunks are then fed directly into the main transformer. This direct input mechanism forces the main transformer to learn the nuances of both perception (understanding the raw input) and thinking (reasoning and generating responses) simultaneously, within a single, cohesive model.

This architectural choice has several profound implications. Firstly, it drastically reduces the model size. By eliminating the need for separate, large modality-specific encoders, Gemma 4 models are up to 99% smaller than comparable multimodal models. For instance, while some open and free AI models like DeepSeek can reach over 1.6 trillion parameters, Gemma 4 offers highly capable models that can run efficiently on consumer hardware like laptops. This smaller footprint translates to lower computational costs, faster inference times, and greater accessibility for developers and researchers.

Secondly, the integration of perception and thinking into a unified model leads to enhanced reasoning capabilities across modalities. The model doesn't just translate inputs; it inherently understands and reasons about them in a deeply interconnected way. This allows Gemma 4 to perform complex tasks that require understanding both visual and textual cues, such as accurately describing objects and actions in a video stream while generating eloquent, contextually relevant prose. The technical report details how Gemma 4 achieves superior performance in tasks requiring agentic reasoning and tool calling, which are critical for developing more sophisticated AI applications.

DeepMind's release of the Gemma 4 Technical Report is also a significant contribution to the open-source AI community. By sharing the architectural details and methodologies, they are enabling other researchers and developers to understand, replicate, and build upon these advancements. This commitment to openness is crucial for accelerating progress in AI, particularly as capabilities increase and the debate around open-source AI intensifies. The Gemma 4 ecosystem is continuously being improved, with ongoing enhancements aimed at making the models even faster and more efficient, further demonstrating the potential of this integrated, multimodal approach.

Open-source AI just took a scary turn #AI #cybersecurity #opensource #AIsafety #technology

Nate B Jones · 2026-08-06

The video highlights a critical concern regarding the increasing prevalence of open-source AI models and their potential to evolve into cyber threats and cyber weapons. The speaker predicts that by the second half of 2026, these models will be ubiquitous on the internet, posing significant risks as they can be leveraged by malicious actors. This shift signifies a new frontier where accessible AI tools can be weaponized, demanding immediate attention to cybersecurity implications.

Are Chinese AI models actually catching up?

Nate B Jones · 2026-08-04

This video argues that Chinese AI models are not catching up to American models despite media narratives. The speaker claims that public benchmarks misrepresent the true state of AI development, as they only reflect released models. He asserts that top American AI labs like Anthropic and OpenAI maintain a significant lead with their internal, unreleased models, and that this gap has remained consistent over the past year.

Anthropic launches Record a Skill for automating workflows #anthropic #recordaskill #automation

AI Honeycove · 2026-08-01

Anthropic's new "Record a Skill" feature in Claude Cowork allows users to teach Claude any workflow by narrating steps while performing them across applications. Claude observes screen activity, clicks, and user narration to understand the underlying logic, then builds a reusable skill. Once recorded, users can invoke these skills with a simple command, and Claude executes the entire workflow, handling tasks like content repurposing, report generation, and data cleanup autonomously, including scheduling for recurring tasks.

NVIDIA is taking on Anthropic?

Theo - t3.gg · 2026-07-31

NVIDIA CEO Jensen Huang made his first Twitter post, sharing an NVIDIA-signed letter that advocates for open-weight AI models built by every company and country to foster innovation and security. The letter subtly criticizes Anthropic, a prominent AI company known for its closed-weight models, by conspicuously omitting Anthropic from a list of over 100 AI companies that supposedly support open models. This move signals a growing divide in the AI community regarding model transparency and accessibility, with NVIDIA positioning itself firmly in the open-model camp while highlighting the perceived closed nature of competitors like Anthropic.

Terraform, VPC, and why you want a tfstate file per env

Charity Majors · 2016-03-30 · 38 min read

TLDR: A single shared Terraform state file across environments is a disaster waiting to happen — one bad change in staging can corrupt state and take down everything, forcing you to manually delete and rebuild your entire infrastructure. The fix is straightforward: maintain separate `tfstate` files per environment so a staging experiment can never poison production state. As a corollary, always hunt for the `panic` in Terraform crash logs (not the error list at the bottom), and never use `-target` on individual modules when state is already dirty, as it silently creates duplicate resources.

datasette 1.0a38

Simon Willison · 2026-08-06 · 1 min read

Datasette 1.0a38 patches a SQL injection vulnerability where users with access to any public table could bypass the permissions system and read data from private tables in the same database instance. This matters because it's a classic broken access control bug — the permissions layer was enforced at the table level but not enforced when raw SQL execution was permitted, a reminder that layered data access controls must account for every query execution path, not just direct table access.