News

Structured outputs for Anthropic Claude models on Amazon Bedrock

Amazon Bedrock now supports structured outputs for all available Anthropic Claude models. The feature works on both bedrock-runtime and bedrock-mantle. Applications can receive JSON that follows a defined schema without custom parsing, retry logic, or post-processing.

Schema rules apply while tokens are generated

Both endpoints offer JSON Schema output formatting and strict tool use to validate tool parameters. Constrained decoding enforces the schema during token generation. It supports a subset of JSON Schema Draft 2020-12. On bedrock-runtime, structured outputs also support batch inference, streaming as typed events, and cross-Region inference. Compiled grammars are cached per account for 24 hours from first access. Changing a schema's structure triggers recompilation, while changing its name or description does not. (AWS Machine Learning)

thinkidiot take: Schema enforcement now happens during token generation, removing the custom parsing and retry logic described in the announcement. I would use strict tool parameters first, where a defined structure is the point of the interface. I would also keep schema structures stable to make use of the 24-hour grammar cache. For my workflows, enforced structure matters more than flexibility in how a model formats its answer.

GenRecs Tokenization: Transforming Ads Recommendation at Meta with Stable Model Vocabulary

Meta's GenRecs Tokenization represents items with tokens from a stable semantic vocabulary instead of dedicated item IDs. Paired with Meta's Generative Ads Model, or GEM, and Adaptive Ranking Model, it delivered a 0.35% lift in advertiser conversions across surfaces. The reported gain comes from that combination of tokenization and models.

New items can connect to what a model already knows

Semantic tokens let models relate new items with no engagement history to familiar items. The stable vocabulary also lets models carry insights across advertisers' campaigns. Its scope includes ads, products, pages, and pixels. It supports both retrieval and ranking across those categories. Meta describes Generative Recommenders as moving from scoring every ad to considering ad content and user preferences together. The approach draws on world knowledge built into large language models to match ads. (Meta Engineering)

thinkidiot take: The 0.35% conversion lift belongs to tokenization paired with GEM and Adaptive Ranking Model. I would keep that combination intact when assessing the result, rather than crediting the vocabulary alone. My first test would focus on items with no engagement history, because semantic tokens explicitly connect them to known items. I find that capability more compelling than the headline lift: it gives the representation a clear job beyond identifying an item.

Mamdani says feds have abdicated on AI, urges New Yorkers to push back

Mayor Zohran Mamdani urged New Yorkers on Monday night to reject the idea that an AI takeover is inevitable. TechCrunch's October 19, 2026 article reports that he spoke with Ilana Glazer at The People's AI Assembly, hosted by the AI Now Institute. He called for regulations that protect the public and put power back in residents' hands. He also argued that public officials should hold AI companies accountable.

Schools and property listings put the argument into practice

Mamdani ordered a ban on AI use in elementary and middle schools the previous month. Over the summer, he introduced a policy addressing AI-generated property listing images. That policy would require landlords, brokers, and real estate platforms to disclose AI-generated photos or videos. His criticism also targeted the federal government's approach to AI. President Trump's Superintelligence Task Force was announced earlier in October. The article describes its focus as largely centered on AI company self-regulation. (TechCrunch)

thinkidiot take: Mamdani has already ordered an AI ban in elementary and middle schools. That makes his position more concrete than a call for accountability alone. I would prioritize the proposed disclosure requirement for property images, because it tells people when the material they are viewing was generated. Disclosure is the more persuasive intervention to me: it addresses a specific use of AI and gives the viewer information.

Embedded AI evaluators won’t be an easy solution, Hugging Face’s policy chief says

Hugging Face chief policy officer Irene Solaiman challenged the idea that embedded AI evaluators offer an easy solution to oversight. She discussed evaluations at TechCrunch Disrupt, according to the October 19, 2026 article. The approach places third-party observers inside AI organizations. It has gained attention over recent months as safety incidents have increased concern about AI harms.

The model is only part of what evaluators must assess

Solaiman identified reproducibility and sufficient model access, including open weights, as important to assessing results. She also said evaluator qualifications need more standardization across tasks. Anthropic hired Accenture as its first evaluator, drawing criticism over the consulting firm's qualifications. Solaiman said a weaker model with a better harness can outperform a frontier model. She described agentic evaluations as difficult and expensive. Fewer evaluations are available for harnesses than for models. (TechCrunch)

thinkidiot take: Harnesses have fewer evaluations available than models, according to Solaiman. I would insist on evaluating the harness alongside the model, given her point that a better harness can put a weaker model ahead. That makes the difficulty and expense of agentic evaluation part of the work, rather than a reason to assess only the model. To me, reproducible results and sufficient access carry more weight than an evaluator's appointment.

The fossil fuel comeback is here

The US is building more fossil fuel infrastructure than it has in a decade, according to a report covered by The Verge. Planned fossil fuel infrastructure rose from 5% of all new capacity in 2024 to 30% by the second quarter of 2026. Electricity demand is climbing faster than it has in more than a decade as more data centers, factories, and electric vehicles connect to the grid. Renewable energy growth is slowing at the same time.

Cheaper generation faces a higher cancellation rate

Fossil fuels still generate 60% of US electricity. Wind and solar together generated more electricity than coal for the first time in 2024, when solar accounted for 81% of added generation capacity. If the trend continues, the report projects fossil fuels reaching 70% of the development pipeline by the end of the decade. Planned fossil capacity would reach 68 gigawatts by 2030, up 45% from plans six months earlier. Renewable energy projects are being canceled at a rate of 4.9%, compared with 2.9% for fossil fuel projects. The report puts average costs for new utility-scale solar and wind at $68 to $69 per megawatt-hour, compared with $90 for new combined-cycle natural gas plants. (The Verge)

thinkidiot take: New solar and wind average $68 to $69 per megawatt-hour in the report, below the $90 figure for new combined-cycle gas plants. I would put those costs beside the cancellation rates when assessing the growing pipeline. Renewable projects have the lower reported generation costs and the higher cancellation rate. That is a poor direction for a grid facing faster demand growth, and I find the shift toward fossil capacity hard to celebrate.

Trending AI Papers

Ranking source: Hugging Face Papers for 2026-10-20.

Self-Evolving Coding Agents via Harness-Policy Optimization

A coding assistant depends on both its model and the software that manages how it works. This paper introduces Self-HPO, a training process that improves these parts in alternating rounds. The model examines its mistakes and suggests changes to its working setup. It then trains within the revised setup to solve more coding tasks.

• Problem: Improving the model alone leaves it learning from whatever experiences its existing setup allows. Improving the setup alone can leave it addressing mistakes the model no longer makes. The two parts need to adapt together.
• New idea: Self-HPO alternates changes to the policy, the model that chooses actions, and the harness, the software that manages its interaction with its environment. The model reviews its failures and proposes harness changes. It then learns under a selected harness through reinforcement learning, training that uses feedback on its actions. Tasks that remain difficult become the material for the next round of harness changes.
• Simple example: Think of a programmer who adjusts their work routine after reviewing mistakes, then practises with that routine. The problems they still struggle with guide the next adjustment.
• Evidence: Across three coding benchmarks, Self-HPO raised Qwen3.5-35B-A3B's average resolve rate from 57.8% to 64.8% and beat approaches that improved only the harness or only the policy. Gains also carried over to settings outside the training distribution. Combining the improved components produced a gain greater than the sum of their separate gains.
• Limitation: The abstract does not report the computing cost of repeated harness searches and model training, so it does not establish how expensive the gains are.
• Why it matters: A coding assistant can solve more tasks when its training and working setup improve together.
• Paper: Self-Evolving Coding Agents via Harness-Policy Optimization

MT Forcing: Towards Stable and Consistent Multi-Turn Image Editing with Autoregressive Diffusion

An image editor should remember earlier requests when you ask for another change. Repeated edits can instead damage details or alter parts you wanted to keep. MT Forcing changes how image-editing models are trained to help them preserve the picture across a sequence of requests.

• Problem: Existing methods can introduce visible defects and lose detail after just a few edits. They can also change untouched areas or undo earlier requests. Each new edit can carry forward the mistakes already present in the image.
• New idea: MT Forcing trains a student model, the editor being improved, with guidance from a teacher model that supplies training targets. For visual quality, the teacher applies all requests so far directly to the original image, avoiding errors carried over from intermediate versions. A second form of guidance uses reference editing sequences, examples of how an image should change as requests accumulate. This teaches the student to respect editing history as well as produce a good-looking result at each step.
• Simple example: Imagine asking an editor to change a shirt's colour, then replace the background. The second edit should keep the new shirt colour and leave the person's face intact.
• Evidence: The abstract reports more stable image quality and better preservation of content across successive edits in extensive experiments. It gives no numerical scores or improvement sizes.
• Limitation: The abstract does not say how long the tested editing sequences were, leaving the method's reliability over longer sessions unclear.
• Why it matters: Preserving earlier edits would let people refine an image without repeatedly repairing changes they never requested.
• Paper: MT Forcing: Towards Stable and Consistent Multi-Turn Image

RoAR-VAE: Improving Latent Diffusability through Robust Prediction

Image generators can work with compressed representations of pictures. Those representations need to retain detail while being easy for the generator to learn from. RoAR-VAE trains this compression system to keep useful information recoverable even when its representation is damaged.

• Problem: A visual tokenizer, which converts images into compressed representations, must support both faithful reconstruction and effective generator training. Earlier work uses guidance about image meaning or trains the tokenizer alongside the generator. The paper finds that parts of one such approach can be removed without sacrificing comparable generation results in a matched comparison.
• New idea: RoAR-VAE trains on corrupted latents, compressed image representations that have been disturbed. Its RobustAlign objective asks the model to recover features describing high-level image content, using features from an already trained model as targets. Its RobustRecon objective asks it to rebuild the clean image, preserving appearance and spatial detail. Together, these training goals make useful image information recoverable despite corruption.
• Simple example: Think of packing a picture into a compact set of notes. Even if some notes get smudged, the remaining information should let someone recover both the overall scene and its small details.
• Evidence: RoAR-VAE improved average generation benchmark performance by 2.19 points over the strongest baseline tokenizer under the same training budget. It also improved small-text reconstruction accuracy by 16.93 percentage points over FLUX.1-VAE.
• Limitation: The abstract does not identify the generation benchmarks or their scoring scales, so the practical size of the 2.19-point gain is hard to assess.
• Why it matters: Better compressed image representations can preserve fine details while making image generators easier to train.
• Paper: RoAR-VAE: Improving Latent Diffusability through Robust

Trending AI Repositories

Ranking source: GitHub Trending.

headroomlabs-ai/headroom

Headroom is a Python project for managing the context sent to AI agents. Its README shows a 55,957-token prompt reduced to 24,340 tokens, with the FATAL line at item 67 preserved byte for byte.

• What it is: Headroom compresses the context sent to AI agents and is available as a library, proxy, and MCP server. The repository claims it reduces token use while preserving answers.
• What it does: Compress tool outputs, logs, files, and RAG chunks before they reach the LLM. 20% fewer tokens for coding agents, 60-95% fewer tokens for JSON, same answers. Library, proxy, MCP server.
• Who it helps: It helps developers building AI agents that work with large prompts. They can reduce the context sent to the model while retaining details such as the error line shown in the README.
• Limitation: The supplied README example shows one preserved error line, not a guarantee that every important detail survives.
• Repository: headroomlabs-ai/headroom

shiyu-coder/Kronos

Kronos is a Python model project focused on financial markets. Its README links to a live demo, giving readers a place to explore the project.

• What it is: Kronos is a Python foundation-model project focused on financial markets.
• What it does: Kronos: A Foundation Model for the Language of Financial Markets
• Who it helps: It is relevant to developers exploring models for financial markets. They can follow the README's links to the live demo and Hugging Face page.
• Limitation: The supplied README excerpt gives no evaluation results or setup requirements.
• Repository: shiyu-coder/Kronos