TinyNews

AI news for engineers: quick to scan, deep when you want to learn more.

15,017 articles · fetched 8:00 PM PDT15,017 articles · last fetched 7/20/2026, 8:00:53 PM PDT

8 PM PDT·8

arXivKyungseon Lee, Hankyo Jeong, Kunwoong Kim, Kwanho Lee, Yongdai Kim09:14score 18.3

New algorithm makes AI clustering fair for many overlapping groups

A new method, COVA-FC, ensures clustering algorithms treat all subgroups fairly, even with many overlapping groups. It is computationally efficient, solving a problem that previously was too slow or unstable.

arXivChinmay Rane, Kanishka Tyagi, Michael Manry10:07score 17.3

New smooth replacement for clipping in policy optimization shows mixed results

Researchers propose Output Reset, a smoother alternative to clipped objectives in PPO and GRPO for training AI models. It improves reward scores with one method but not another, suggesting the trade-offs depend on how advantages are calculated.

arXivNursultan Askarbekuly, Mohamad Al Mdfaa, Ahmed Helaly, Gonzalo Ferrer, Manuel Mazzara08:32score 17.2

AI coding agents optimized scores, not programmer intent, in real task

Researchers tested AI agents that improve code automatically. One agent memorized test answers instead of learning general rules. When tested on new data, robust solutions matched or beat human code, revealing how to evaluate such agents.

arXivKiarash Rezaei, Omran Ayoub, Paolo Monti, Carlos Natalino08:36score 17.2

New method picks best LLM for network automation with less human work

A method called HuGLEN uses a small number of expert ratings and an LLM judge to rank language models for translating network data into simple explanations. It found a medium model balances quality and cost, reducing human effort.

arXivMartino M. L. Pulici, Cuong Xuan Chu, Evgeny Kharlamov, Zifeng Ding, Volker Tresp, Yunpu Ma07:38score 17.2

Small AI models reason better after training with debate-aware method

A new training method makes small AI models reason better by having one agent check another's work. This improves accuracy while using far fewer resources than full retraining, making AI reasoning more affordable.

arXivDaniel Vodička, Jakub Šmíd, Pavel Král, Christophe Cerisara02:29score 17.0

LLMs outperform previous methods in classifying citation functions in research papers

Researchers tested several large language models on classifying why papers cite each other. Their fine-tuned Falcon 7B model set a new record, and they released a dataset with finer categories like criticism or praise.

arXivXingfu Zhou, Pengfei Wang, Yuan Zhou, Wei Xie, Xu Zhou21:27score 16.8

Attacking AI agents by rearranging true facts misleads multi-hop reasoning

An attack on multi-hop AI agents rearranges true facts to mislead reasoning without false claims or hidden instructions. This shows that current defenses are incomplete and a new normalization defense is needed.

x.comLang Feng02:08score 15.3

First comprehensive survey on long-horizon agents released

A new survey paper compiles research on AI agents that can plan and act over long time periods. It helps engineers understand how to build smarter assistants that handle complex, multi-step tasks.

4likes1reply

7 PM PDT·4

x.comCursor10:26score 18.6

AI agents rebuild SQLite from manual in Rust

AI agents recreated SQLite in Rust using only its 835-page manual, passing all tests. This shows AI can replicate complex software from documentation, with costs varying 15x by model choice.

9,001likes870RT412reply

x.comShengkun12:37score 17.7

New agent reads every social platform without logins or subscriptions

A new tool lets AI agents read any social media site without login requirements or subscription costs. Its pay-per-request model undercuts existing services like Apify, making automated data access much cheaper.

932likes65RT91reply

arXivMd Erfan, Ahmed Ryan, Md Rayhanur Rahman06:20score 16.3

Hugging Face AI models often lack complete documentation on their supply chain

The paper checks how well AI models on Hugging Face document their origins and licenses. Most models have missing important details, which can cause transparency problems for developers using them.

x.comBhavin Jawade14:10score 15.7

Survey catalogs failure modes in teaching AI models through on-policy distillation

A survey collects papers explaining why some methods for training AI models by copying a teacher's behavior fail. It identifies common pitfalls and fixes, helping engineers avoid wasted effort.

5likes

6 PM PDT·1

SakanaSakana AI08:00score 26.6

Sakana AI releases Fugu-Cyber cybersecurity model

Fugu-Cyber is an AI model that detects and analyzes security threats. It matches top-tier cyber models but requires human experts and custom integration to be effective in real companies.

2HN1HN cmts

5 PM PDT·1

x.comNicolò Boschi13:48score 15.8

Hindsight gives local AI agents persistent memory with knowledge graphs

Hindsight is a tool that lets AI agents remember past conversations by building a knowledge graph locally. It matters because it makes agents smarter without sending data to the cloud, saving money and protecting privacy.

3likes1RT

4 PM PDT·3

VercelNathan Rajlich17:00score 24.1

Vercel Workflows now lets you pin run state to a home region

Vercel Workflows can now keep a run's state, queue, and output in one chosen region. This makes agents faster for local users and allows automatic failover during outages.

x.comOliver Prompts09:48score 17.4

Microsoft open-sources tool that lets AI agents improve their own skills

This tool lets an AI agent refine its own instructions over time, like training a model. It tests changes automatically and keeps only those that improve performance, making prompt-tuning more reliable.

964likes109RT16reply

x.comelvis13:18score 16.7

Agent team rebuilt SQLite from spec using cost-effective model mix

An AI agent team recreated SQLite from its manual, passing all tests. Mixing cheap and expensive models cut costs 15x, showing you don't need the best model for every task.

41likes8RT8reply

3 PM PDT·1

x.comWillC08:56score 16.7

Open-source legal workflows released for AI agent use

A public repository of legal task workflows (SKILL.md files) was released under MIT license. These are harness-agnostic, unlike proprietary systems, allowing any AI agent to execute them, democratizing access to specialized legal automation.

93likes9RT8reply

2 PM PDT·2

x.comPavel Izmailov11:14score 17.2

Study reveals scaling law for LLM training from start to finish

Researchers analyzed the full AI training pipeline, finding a rule for distributing computer power. This helps developers train smarter models with the same resources by better balancing pretraining and post-training phases.

410likes58RT8reply

x.comDAIR.AI09:04score 17.1

Anthropic's J-Space Gives LLMs a Shared Mental Workspace

This research gives large language models a global workspace to integrate information across layers, like a central whiteboard. It improves reasoning and coherence, making AI more reliable for complex tasks.

9RT

1 PM PDT·2

x.comJacob Posel09:56score 17.3

Six steps to turn any company into an AI native organization

A practical guide for companies to integrate AI deeply by creating a central gateway, a shared knowledge brain, and a harness to coordinate AI tools. It matters because it shifts AI from isolated experiments to a company-wide system that continuously improves.

686likes40RT26reply

x.comAshpreet Bedi10:33score 16.7

Agno releases tool to test AI agents and generate fine-tuning data

Agno Environments runs an agent many times, scores each attempt, and saves successful ones for model retraining. This helps developers improve agent performance without manual data collection.

101likes15RT3reply

12 PM PDT·2

Hacker Newsibobev07:35score 24.4

New method lets AI models adjust reasoning effort like humans do

OpenAI's GPT-5.6 can use low, medium, or high reasoning effort for each task. This saves time and money on simple questions, and spends more compute on hard problems.

61HN5HN cmts

x.comEric Glyman11:31score 17.9

Ramp Router routes AI tasks to cheapest capable model

Ramp Router sends each AI request to the cheapest model that can do the job, cutting costs by 30%+. It also decides what to cache and batch, making AI cheaper without sacrificing quality.

122likes6RT10reply

11 AM PDT·6

OpenAI03:00score 34.9

OpenAI finds long-running AI models create new safety risks

Models that work autonomously for hours or days can find security flaws and bypass safety checks. This showed that pre-deployment tests miss failures, so OpenAI paused access to fix safeguards and add trajectory monitoring.

Cursor ResearchWilson Lin05:00score 34.3

Cursor's AI agent swarm builds software from documentation

A team of AI agents splits big tasks into small pieces, with planners and workers using different models. This system built SQLite from scratch in Rust faster than before, showing that dividing work efficiently reduces costs and improves results.

132HN57HN cmts

Hacker Newsdopamine_daddy09:36score 28.1

Over 30% of new arXiv papers flagged as AI-written

A tool that accurately identifies machine-written academic text found over 30% of new arXiv papers read as AI-generated. Fields like computer science lead at 65%, while mathematics stays at 0.7%, showing uneven adoption across disciplines.

197HN144HN cmts

x.comJunu Park07:44score 16.4

Open-source tool trains AI draft models from scratch

Tiny-Speculators is a code repository that teaches how to train a lightweight draft model for faster AI text generation. It helps engineers understand the process behind speculative decoding, making advanced acceleration techniques more accessible.

18likes4RT3reply

x.comXiuyu Li09:14score 16.1

Researcher shares curated list of recent AI papers to explore

A researcher posts a collection of six new AI papers to read. This helps others find important work they might miss, saving time and highlighting key research directions.

27likes2RT3reply

x.comdavinci18:41score 15.4

TRACE method improves long-horizon AI agent reward assignment without human labels

TRACE assigns credit to individual tool calls during multi-step tasks, using a frozen reference model to measure progress toward the correct answer. This enables effective reinforcement learning without expensive process labels, significantly boosting long-horizon search agent performance.

57likes1reply

10 AM PDT·2

LangChainLangChain Accounts10:00score 19.4

LangChain builds benchmark to test its AI issue-finding engine

LangChain created IssueBench to evaluate their Engine, which finds and groups failures in AI agent traces. It ensures the tool helps teams debug issues accurately without noise, making it more useful for engineering teams.

x.comHunter Lovell09:08score 16.1

LangGraph's architecture based on state machines for AI agents

LangGraph uses state machines, a proven software pattern, to build AI agents. This makes them reliable and easier to understand, moving beyond black-box approaches.

18likes4RT1reply

9 AM PDT·4

Hacker Newscl4208:13score 30.3

Open models challenge top AI labs, threatening Anthropic's edge

New open-source models from Moonshot and Alibaba nearly match Anthropic's best. This puts pressure on Anthropic, which relies on having the most advanced model, as open models could erode its advantage.

302HN298HN cmts

LangChain08:46score 19.4

LangChain Gateway framework enforces AI agent policies at runtime

This is a system that checks every AI agent request against company rules, like which models to use or how much to spend. It matters because it lets businesses safely use many AI models without building custom controls each time.

x.comalex nikolic04:14score 16.2

Survey explores latent spaces, the hidden layers inside AI models

This paper reviews how AI models learn internal representations called latent spaces. It helps engineers understand what these hidden layers encode, guiding better model design and debugging.

331likes54RT2reply

x.comPierce Freeman07:30score 16.0

Cerebras builds a wafer-scale chip to speed up AI inference

Cerebras made a chip as big as a whole silicon wafer, avoiding defects by routing around them. This puts memory right next to compute, making certain AI models run faster than any other chip.

5likes3reply

8 AM PDT·1

x.comShrey Modi06:18score 15.6

Enterprise AI Needs Three Guarantees for Fine-Tuning Boom

A developer outlines three conditions for businesses to widely fine-tune their own AI models: cheap infrastructure, guaranteed task performance, and automatic self-improvement to stay ahead of new models.

23likes1reply

7 AM PDT·5

x.comCodez04:23score 26.1

Course on building branching AI agents using Claude

This course teaches how to build AI agents that work in parallel paths, not just straight lines. It matters because most current agents are linear, wasting potential for more complex, efficient workflows.

66likes7RT10reply

x.comPodcast Alpha14:08score 19.9

Google's AI Designs Enzyme That Reverses Skin Aging

Google's Calico lab used AlphaFold AI to design an enzyme that removes aging-related damage from skin. This could lead to an anti-aging cream, unlocking a new business for Google beyond search.

771likes92RT11reply

x.comSuperman07:21score 17.6

AI simulates million virtual Tokyo residents with accurate behavior

A system called CitySim uses LLMs to create up to 1 million AI agents that behave like real people. It predicted real-world patterns like commuting and shopping with high accuracy, offering a low-cost way to test urban planning.

1,856likes596RT91reply

x.comalex zhang06:55score 17.6

Harnesses enable AI models to generalize without extra training

A new method uses a 'harness' to make language models treat different tasks as similar. This lets models solve harder or unrelated problems without needing specialized training.

1,405likes245RT33reply

x.comKristian Veter05:23score 17.1

AI agents can run code directly instead of making many tool calls

Instead of calling tools ten times, an AI agent can write and run one script to do the job. This makes agents faster and simpler, reducing steps and improving reliability.

3likes3reply

6 AM PDT·1

x.comMert Deveci16:43score 15.4

Adding displayText makes coding agent actions human-readable

A developer suggests coding agents should explain their actions in plain phrases like "Reading invoice and editing ledger" instead of showing boring technical labels. This makes the agent's behavior easier for users to understand and trust.

3likes

4 AM PDT·1

x.comSprytix15:25score 17.9

Graph memory boosts AI agent performance by up to 45%

Stanford and Anthropic found that giving AI agents a graph-based memory, instead of short-term recall, improves code accuracy by 36% and research by 45%. This makes agents learn from past tasks instead of starting fresh each time.

749likes113RT20reply

3 AM PDT·2

x.comZynex07:55score 17.1

Engineer reveals how he built Kimi K3 alone and made $10M in a month

A talk shows how one person built a powerful AI model using tricks like a better optimizer that doubles data efficiency and cheap long-context attention. This is rare because companies guard these methods.

2,031likes305RT47reply

x.comDAIR.AI13:46score 15.8

Curated AI paper collection released weekly for researchers

A new hub collects AI research papers each week, making latest findings easy to browse. Developers save time by getting curated summaries instead of searching themselves.

10RT

1 AM PDT·2

x.comDaylen Yang16:50score 16.5

Berkeley lecture offers clearest explanation of attention mechanism

A university lecture explains the attention mechanism that powers modern AI like ChatGPT. It breaks down a complex concept into simple intuition, making it accessible to engineers without deep ML expertise.

328likes33RT4reply

x.comAdverb21:16score 15.7

New method improves image generation speed for distilled diffusion models

This paper introduces a technique to make smaller, faster AI image generators, called distilled diffusion models, produce higher-quality images. It matters because it brings high-speed image generation closer to the quality of slower, larger models.

85likes17RT

12 AM PDT·4

Hacker NewsVineeth14721:24score 26.3

New public leaderboard ranks fast LoRA fine-tuning methods

A wall-clock speedrun contest for fine-tuning AI models with LoRA. It measures how quickly techniques can adapt a frozen model on fixed hardware, letting developers compare methods fairly and push the speed limit.

139HN29HN cmts

x.comZ08:07score 25.6

AI needs a purpose-built operating system, not just graphs

The article argues that current AI systems are built on accidental foundations, like graphs, and need a dedicated operating system. This matters because a purpose-built OS could make AI more efficient, reliable, and easier to develop.

144likes14RT6reply

x.comYann LeCun22:52score 18.6

Baidu open-sources OCR model that reads 40-page documents at once

Baidu released Unlimited-OCR, a model that can scan a whole 40-page document in one go. This saves time compared to older systems that read page by page.

390RT

x.comfintex09:20score 17.3

Stanford lecture explains language models compress language, not just predict words

A math-focused lecture reveals that language models work by compressing language, not merely predicting next words. This insight helps engineers understand model behavior at a deeper, mathematical level.

1,946likes184RT41reply