News

Bring more intelligence to everyday work with GPT-6 Sol and GPT-6 Luna on Amazon Bedrock

AWS has made GPT-6 Sol and GPT-6 Luna generally available on Amazon Bedrock. Both models have significantly lower API pricing than their GPT-5.6 predecessors. Bedrock users can now choose between the new models to match intelligence and efficiency to each workload.

Adjustable reasoning and a measured drop in factual errors

GPT-6 Sol can implement features, debug issues, refactor code, and review it. It also supports data analysis. Its capabilities extend to completing multistep processes across tools and applications. On an internal OpenAI factuality evaluation, Sol made approximately half as many factual mistakes as GPT-5.6 Sol. GPT-6 Luna supports document extraction, summarization, classification, and focused question answering. Luna also lets users adjust reasoning effort per request to balance quality, responsiveness, and cost. (AWS Machine Learning)

thinkidiot take: Both models now cost significantly less through the API than their GPT-5.6 predecessors. I would start with Luna on document extraction and classification, then vary reasoning effort against the quality each task requires. That makes the quality and cost tradeoff a decision for each request. For routine work, I value that control more than Sol's broader list of capabilities.

TechCrunch Founder Summit’s agenda revealed: Unlock fundraising, hiring, and AI insights in Boston on November 4

TechCrunch has revealed the agenda for its Founder Summit on November 4 at Boston's SoWa Power Station. The one-day event focuses on building a startup. Founders can now see the planned discussions on fundraising, leadership, and building an AI-native company.

The investor gets evaluated too

Underscore partner Brian Devaney will discuss investor expectations and mistakes that cost founders leverage. HubSpot co-founder and Sequoia partner Brian Halligan will cover how founders evolve as CEOs through company growth. 7AI co-founder and CEO Lior Div will examine team structure, product development, operations, and go-to-market strategy. Cogent Security co-founder and CEO Vineet Edupuganti will discuss evaluating investors after raising an $11 million round. TDK Ventures investment director Tina Tosukhowong will explain its “King of the Hill” framework. That framework evaluates economics, scalability, and commercial timing. (TechCrunch)

thinkidiot take: The agenda includes investor evaluation from a founder who has raised an $11 million round. I would prioritize Edupuganti's session alongside Tosukhowong's framework for economics, scalability, and commercial timing. Together, those topics give me a concrete set of questions to take into a fundraising discussion. I judge that more useful than advice focused only on getting an investor to say yes.

Snorkel AI triples valuation to $3.5B as demand for AI training data booms

Snorkel AI has raised a $350 million Series E led by Insight Partners and S32 at a $3.5 billion valuation. The funding will support the seven-year-old startup's data-as-a-service approach. Its previous round was a $100 million Series D at a $1.3 billion valuation 17 months earlier.

Completed datasets replace labeling software

Snorkel says its annualized revenue run rate reached $375 million, an eighteenfold increase over the last 12 months. Last year, it shifted from data-labeling automation software to selling completed datasets. It uses software and models to generate synthetic data alongside subject matter experts. The company also sells reinforcement learning environments. Snorkel says payments to human experts are accounted for in cost of goods sold. It launched commercially in 2019 after four years of research by Alex Ratner and his team at a Stanford AI lab. (TechCrunch)

thinkidiot take: Snorkel now sells completed datasets instead of data-labeling automation software. I would evaluate that offering by examining the delivered data and reinforcement learning environments. Human expert payments sit in cost of goods sold, so I would also want to understand the economics behind the reported $375 million annualized revenue run rate. For me, the change in what customers buy matters more than the $3.5 billion valuation.

Better prompt caching for GPT-6

OpenAI has announced improvements to prompt caching for GPT-6. Cache hit rates are higher. The announcement also introduces ways for developers to inspect and control caching.

Diagnostics and explicit breakpoints enter the workflow

The update includes new diagnostics and explicit breakpoints, along with controls to reduce latency and costs. (OpenAI)

thinkidiot take: GPT-6's caching controls reduce latency and costs. I would use the new diagnostics while trying explicit breakpoints in my prompts. My test would focus on those two stated benefits. I value that practical control more than a higher cache hit rate on its own.

Trending AI Papers

Ranking source: Hugging Face Papers for 2026-09-23.

RULER: Instance-aware Rubric Rewards for SVG Generation

Teaching a model to draw from written instructions requires a way to judge its drawings. RULER uses a checklist tailored to each request to assess images made with SVG, a format that describes graphics in code. It turns those assessments into feedback for training. The aim is to reward drawings that meet the request across content, appearance and style.

• Problem: There is no single correct drawing for an open-ended request. Existing scoring tools built around natural images judge stylized vector graphics poorly. Training models to maximize those scores can teach them to exploit the scoring system.
• New idea: RULER turns each written request into a rubric, a checklist of six criteria covering requested content, visual qualities and style. A vision-language model, which processes both images and text, checks the rendered drawing against each criterion. The results are combined into a weighted reward, so criteria can contribute different amounts to the training signal. Reinforcement learning uses that reward to improve the drawing model without example SVG answers or human preference labels.
• Simple example: Think of judging a drawing assignment with a checklist written for that assignment. Checking whether the drawing includes the requested content and follows the requested style gives more useful feedback than a single general beauty score.
• Evidence: Rubric scores rose from 0.432 to 0.693 on MMSVG-Illustration and from 0.395 to 0.683 on MMSVG-Icon. RULER outperformed dedicated SVG models and matched the much larger DeepSeek-V3. The rubric-based judge also agreed with human judgments much better than the scalar metrics tested.
• Limitation: The abstract reports better agreement with human judgments, but does not establish how reliably the rubric judge resists exploitation during training.
• Why it matters: Better feedback can help drawing models learn to follow a request more closely.
• Paper: RULER: Instance-aware Rubric Rewards for SVG Generation

Circuit Hypernetworks for Quantum-Augmented Diffusion Language Models

HyperQ explores a way to adapt a language model while leaving its existing weights untouched. It adds small trainable components that use quantum circuit calculations to adjust how the model handles pieces of text. Those calculations can be evaluated exactly on a classical computer. The work aims to make larger circuits practical inside a language model.

• Problem: Quantum circuits offer a way to change how a language model processes individual tokens, the pieces of text it works with. But evaluating larger circuits inside a large model can demand too much computation.
• New idea: HyperQ keeps the original language model fixed and trains added branches, small modules that contribute adjustments to its internal calculations. Each branch reads the model's internal representation of a token and uses a hypernetwork, a network that produces another computation's settings, to configure a quantum circuit for that token. The circuit's measured values are added back to the model's internal representation. An exact classical formula evaluates the required averages with a cost that grows linearly with the number of qubits, the circuit's quantum information units.
• Simple example: Think of keeping a text-processing machine intact and attaching an adjustable helper at each stage. The helper changes its settings for each piece of text and feeds a correction back into the machine.
• Evidence: HyperQ trained circuits from 16 to 64 qubits inside a 1.1-billion-parameter model. Increasing circuit width raised the average benchmark score from 47.65 to 54.30. At 64 qubits, it beat the original model by 4.71 points and a version using low-rank adaptation by 3.67 points. HyperQ used 20,000 prompt-response pairs for fine-tuning, compared with 200,000 for the classical baselines.
• Limitation: The results do not establish an advantage from running on quantum hardware: the required circuit values were evaluated using an exact classical expression.
• Why it matters: This offers a practical way to test quantum circuit calculations as trainable additions to a language model.
• Paper: Circuit Hypernetworks for Quantum-Augmented Diffusion

GAE: Learning a Geometry-Native Latent Space for 3D-Consistent World Generation

A generated scene can look convincing in individual images yet fail to make sense as a three-dimensional place. GAE changes the compact internal description a generator uses to build that scene. It draws on a model that already extracts spatial structure from images. The aim is to make visual detail and scene geometry work together during generation.

• Problem: Visual generators can make realistic frames without keeping the underlying scene consistent across views. Their internal descriptions usually emphasize appearance, leaving spatial structure poorly represented. Adding geometry as another output does not address that underlying representation choice.
• New idea: GAE is an autoencoder, a model that compresses information into an internal representation and reconstructs outputs from it. It builds that compact representation from features produced by a geometry model, which extracts spatial structure from images. The same representation can be decoded into appearance, depth, camera information and point maps, which describe locations in three-dimensional space. A generator then works with this shared representation to produce scenes.
• Simple example: Think of drawing views of a room from a shared model of its layout. That gives each view a common spatial reference, instead of relying only on how convincing each separate drawing looks.
• Evidence: With the generator and training procedure held fixed, using GAE reduced FVD, the reported visual-quality measure, by 12.7% on RealEstate10K and 23.1% on DL3DV. Camera-trajectory error was halved on RealEstate10K. The comparisons found improvements in both visual quality and independently measured 3D consistency.
• Limitation: The abstract reports results on RealEstate10K and DL3DV, but does not establish how broadly the gains transfer to other kinds of scenes.
• Why it matters: A shared spatial representation can help generated views fit together as a coherent place.
• Paper: GAE: Learning a Geometry-Native Latent Space for

Trending AI Repositories

Ranking source: GitHub Trending.

dream-num/univer

Univer is an embeddable Office SDK with a plugin architecture and Canvas-based rendering. Its formula engine and shared Facade API give developers building blocks for productivity features in the browser and on Node.js.

• What it is: This TypeScript SDK lets developers add customizable Office features to their applications, with PDF support coming soon.
• What it does: The Office Harness for AI Agents , Spreadsheets, Docs, Slides, Canvas, Relational Tables, and PDF in one runtime.
• Who it helps: It helps developers embedding Office features in their applications. They can customize those features through plugins and use one Facade API across browser and Node.js environments.
• Limitation: PDF support is listed as coming soon.
• Repository: dream-num/univer

davila7/claude-code-templates

Claude Code Templates is a developer tool in the Claude Code ecosystem. Its README links to an npm package and an MIT license, giving readers a starting point for checking distribution and reuse terms.

• What it is: The repository is listed as Python, and its README points to a package on npm.
• What it does: CLI tool for configuring and monitoring Claude Code
• Who it helps: It helps Claude Code users working on their setup. They can use it to configure Claude Code and monitor it.
• Limitation: The supplied README excerpt contains badges and links, but no setup instructions or feature details.
• Repository: davila7/claude-code-templates

google/ax

AX is a declarative orchestrator for agentic tasks. It handles sandboxing, workspace setup and network boundaries from task specifications, bringing those execution concerns into one system.

• What it is: This Go project sits in the execution layer for agentic tasks, using workspace and gateway specifications to prepare and constrain their environment.
• What it does: Google's open agentic orchestration runtime
• Who it helps: It helps developers running agentic tasks at scale. They can declare workspace and gateway requirements and have AX set up the workspace, sandbox the task and fence its network.
• Limitation: Core concepts, protocols and specifications are still being refined, with major breaking changes likely before a stable release.
• Repository: google/ax