News from the AI & ML world

DeeperML - #llm

@www.marktechpost.com //

Mistral AI Releases Improved Open Source Model Mistral 3.2 - Mistral AI has released Mistral Small 3.2, an updated version of its open-source model that improves instruction following and function calling and complies with EU regulations like GDPR and the EU AI Act.
Mistral AI has released Mistral Small 3.2, an updated version of its open-source model, Mistral-Small-3.2-24B-Instruct-2506, building upon the earlier Mistral-Small-3.1-24B-Instruct-2503. This update focuses on enhancing the model’s overall reliability and efficiency, particularly in handling complex instructions, minimizing repetitive outputs, and maintaining stability during function-calling scenarios. The improvements aim to refine specific behaviors such as instruction following, output stability, and function calling robustness without altering the core architecture.

A significant enhancement in Mistral Small 3.2 is its improved accuracy in executing precise instructions. Benchmark scores reflect this improvement, with the model achieving 65.33% accuracy on the Wildbench v2 instruction test, up from 55.6% for its predecessor. Performance on the challenging Arena Hard v2 test nearly doubled, increasing from 19.56% to 43.1%, demonstrating an enhanced ability to understand and execute intricate commands accurately. Internally, Mistral’s accuracy rose from 82.75% in Small 3.1 to 84.78% in Small 3.2.

Mistral Small 3.2 also addresses the issue of repetitive errors by significantly reducing instances of infinite or repetitive output, a common problem in extended conversational scenarios. Internal evaluations show a decrease in infinite generation errors by nearly half, from 2.11% in Small 3.1 to 1.29%. The updated model also demonstrates enhanced capability in calling functions, making it more suitable for automation tasks. Additionally, Mistral AI emphasizes its compliance with EU regulations like GDPR and the EU AI Act, making it an appealing choice for developers in the region.

Recommended read:
References :
  • Simon Willison: Blogged too much today and had to send it all out in a newsletter - it's a pretty fun one, covering Gemini 2.5 and Mistral Small 3.2 and the fact that most LLMs will absolutely try and murder you given the chance (and a suitably contrived scenario)
  • www.marktechpost.com: Mistral AI Releases Mistral Small 3.2: Enhanced Instruction Following, Reduced Repetition, and Stronger Function Calling for AI Integration
  • AI News | VentureBeat: Mistral just updated its open source Small model from 3.1 to 3.2: here’s why
  • Simon Willison: Mistral Small 3.2 was released today - I used a 15GB quantized model from Hugging Face (via Ollama) running on my Mac and got it to draw me a pretty decent SVG of a pelican riding a bicycle model (considering the model size)
@www.microsoft.com //

Microsoft Advances AI Reasoning and Cloud Services - Microsoft is enhancing reasoning in language models with AI for weather forecasting and rolling out sovereign cloud for Europe.
Microsoft is making significant advancements in artificial intelligence, focusing on improved reasoning in language models and enhanced weather forecasting capabilities. New methods are being developed to boost reasoning in both small and large language models, combining symbolic logic, mathematical rigor, and adaptive planning. These techniques are designed to enable AI models to tackle complex, real-world problems across various fields, potentially transforming AI into a more reliable partner in domains like scientific research and healthcare.

A new AI model named Aurora, developed by Microsoft, can forecast hurricanes and sandstorms up to 5,000 times faster than conventional weather models powered by supercomputers. Aurora outperformed existing systems in predicting weather conditions over a 14-day period in 91% of cases. The model is trained on over 1 million hours of global atmospheric data, including weather station readings, satellite images, and radar measurements, representing one of the largest datasets used to train a weather AI model.

To address the growing demand for data control in Europe, Microsoft is expanding its Sovereign Cloud offerings. This includes solutions that ensure European data remains within Europe, handled exclusively by Microsoft employees based in the region. The Sovereign Public Cloud offers tools and options for customer-controlled encryption and simplified configurations, providing organizations in Europe with greater control over their data. The cloud is offered across all existing European data center regions.

Recommended read:
References :
@www.marktechpost.com //

Apple Research Questioning LLMs Ability in Reasoning - Apple research challenges the reasoning capabilities of Large Reasoning Models (LRMs), suggesting they struggle with basic reasoning tasks, sparking debate within the AI community where Google is on the other side of the debate.
Apple researchers are challenging the perceived reasoning capabilities of Large Reasoning Models (LRMs), sparking debate within the AI community. A recent paper from Apple, titled "The Illusion of Thinking," suggests that these models, which generate intermediate thinking steps like Chain-of-Thought reasoning, struggle with fundamental reasoning tasks. The research indicates that current evaluation methods relying on math and code benchmarks are insufficient, as they often suffer from data contamination and fail to assess the structure or quality of the reasoning process.

To address these shortcomings, Apple researchers introduced controllable puzzle environments, including the Tower of Hanoi, River Crossing, Checker Jumping, and Blocks World, allowing for precise manipulation of problem complexity. These puzzles require diverse reasoning abilities, such as constraint satisfaction and sequential planning, and are free from data contamination. The Apple paper concluded that state-of-the-art LRMs ultimately fail to develop generalizable problem-solving capabilities, with accuracy collapsing to zero beyond certain complexities across different environments.

However, the Apple research has faced criticism. Experts, like Professor Seok Joon Kwon, argue that Apple's lack of high-performance hardware, such as a large GPU-based cluster comparable to those operated by Google or Microsoft, could be a factor in their findings. Some argue that the models perform better on familiar puzzles, suggesting that their success may be linked to training exposure rather than genuine problem-solving skills. Others, such as Alex Lawsen and "C. Opus," argue that the Apple researchers' results don't support claims about fundamental reasoning limitations, but rather highlight engineering challenges related to token limits and evaluation methods.

Recommended read:
References :
  • TheSequence: The Sequence Research #663: The Illusion of Thinking, Inside the Most Controversial AI Paper of Recent Weeks
  • chatgptiseatingtheworld.com: Research: Did Apple researchers overstate “The Illusion of Thinking†in reasoning models. Opus, Lawsen think so.
  • www.marktechpost.com: Apple Researchers Reveal Structural Failures in Large Reasoning Models Using Puzzle-Based Evaluation
  • arstechnica.com: New Apple study challenges whether AI models truly “reason†through problems
  • 9to5Mac: New paper pushes back on Apple’s LLM ‘reasoning collapse’ study
nftjedi@chatgptiseatingtheworld.com //

Apple Research Paper on AI Reasoning Faces Criticism - Apple researchers published a paper questioning the reasoning abilities of LLMs, arguing they rely on pattern matching rather than true reasoning, but critics argue that the experiments were unfairly designed.
Apple researchers recently published a study titled "The Illusion of Thinking," suggesting that advanced language models (LLMs) struggle with true reasoning, relying instead on pattern matching. The study presented findings based on tasks like the Tower of Hanoi puzzle, where models purportedly failed when complexity increased, leading to the conclusion that these models possess limited problem-solving abilities. However, these conclusions are now under scrutiny, with critics arguing the experiments were not fairly designed.

Alex Lawsen of Open Philanthropy has published a counter-study challenging the foundations of Apple's claims. Lawsen argues that models like Claude, Gemini, and OpenAI's latest systems weren't failing due to cognitive limits, but rather because the evaluation methods didn't account for key technical constraints. One issue raised was that models were often cut off from providing full answers because they neared their maximum token limit, a built-in cap on output text, which Apple's evaluation counted as a reasoning failure rather than a practical limitation.

Another point of contention involved the River Crossing test, where models faced unsolvable problem setups. When the models correctly identified the tasks as impossible and refused to attempt them, they were still marked wrong. Furthermore, the evaluation system strictly judged outputs against exhaustive solutions, failing to credit models for partial but correct answers, pattern recognition, or strategic shortcuts. To illustrate, Lawsen demonstrated that when models were instructed to write a program to solve the Hanoi puzzle, they delivered accurate, scalable solutions even with 15 disks, contradicting Apple's assertion of limitations.

Recommended read:
References :
  • chatgptiseatingtheworld.com: Research: Did Apple researchers overstate “The Illusion of Thinking†in reasoning models. Opus, Lawsen think so.
  • Digital Information World: Apple’s AI Critique Faces Pushback Over Flawed Testing Methods
  • NextBigFuture.com: Apple Researcher Claims Illusion of AI Thinking Versus OpenAI Solving Ten Disk Puzzle
  • Bernard Marr: Beyond The Hype: What Apple's AI Warning Means For Business Leaders
Kristin Sestito@hiddenlayer.com //

TokenBreak Attack Bypasses AI Moderation using Character Manipulation - The TokenBreak attack bypasses AI text classification models by manipulating tokens in the input text, inducing false negatives and leaving systems vulnerable.
Cybersecurity researchers have recently unveiled a novel attack, dubbed TokenBreak, that exploits vulnerabilities in the tokenization process of large language models (LLMs). This technique allows malicious actors to bypass safety and content moderation guardrails with minimal alterations to text input. By manipulating individual characters, attackers can induce false negatives in text classification models, effectively evading detection mechanisms designed to prevent harmful activities like prompt injection, spam, and the dissemination of toxic content. The TokenBreak attack highlights a critical flaw in AI security, emphasizing the need for more robust defenses against such exploitation.

The TokenBreak attack specifically targets the way models tokenize text, the process of breaking down raw text into smaller units or tokens. HiddenLayer researchers discovered that models using Byte Pair Encoding (BPE) or WordPiece tokenization strategies are particularly vulnerable. By adding subtle alterations, such as adding an extra letter to a word like changing "instructions" to "finstructions", the meaning of the text is still understood. This manipulation causes different tokenizers to split the text in unexpected ways, effectively fooling the AI's detection mechanisms. The fact that the altered text remains understandable underscores the potential for attackers to inject malicious prompts and bypass intended safeguards.

To mitigate the risks associated with the TokenBreak attack, experts recommend several strategies. Selecting models that use Unigram tokenizers, which have demonstrated greater resilience to this type of manipulation, is crucial. Additionally, organizations should ensure tokenization and model logic alignment and implement misclassification logging to better detect and respond to potential attacks. Understanding the underlying protection model's family and its tokenization strategy is also critical. The TokenBreak attack serves as a reminder of the ever-evolving landscape of AI security and the importance of proactive measures to protect against emerging threats.

Recommended read:
References :
  • Security Risk Advisors: TokenBreak attack bypasses AI text filters by manipulating tokens. BERT/RoBERTa vulnerable, DeBERTa resistant. #AISecuority #LLM #PromptInjection The post appeared first on .
  • The Hacker News: Cybersecurity researchers have discovered a novel attack technique called TokenBreak that can be used to bypass a large language model's (LLM) safety and content moderation guardrails with just a single character change.
  • www.scworld.com: Researchers detail how malicious actors could exploit the novel TokenBreak attack technique to compromise large language models' tokenization strategy and evade implemented safety and content moderation protections
  • hiddenlayer.com: New TokenBreak Attack Bypasses AI Moderation with Single-Character Text Changes
Carl Franzen@AI News | VentureBeat //

Mistral AI Launches Magistral Reasoning Models Openly - Mistral AI launched its first reasoning model, Magistral, available in both large and small Apache 2.0 versions, and introduced Mistral Agents API’s Handoffs feature for smart, multi-agent workflows.
Mistral AI has launched its first reasoning model, Magistral, signaling a commitment to open-source AI development. The Magistral family features two models: Magistral Small, a 24-billion parameter model available with open weights under the Apache 2.0 license, and Magistral Medium, a proprietary model accessible through an API. This dual release strategy aims to cater to both enterprise clients seeking advanced reasoning capabilities and the broader AI community interested in open-source innovation.

Mistral's decision to release Magistral Small under the permissive Apache 2.0 license marks a significant return to its open-source roots. The license allows for the free use, modification, and distribution of the model's source code, even for commercial purposes. This empowers startups and established companies to build and deploy their own applications on top of Mistral’s latest reasoning architecture, without the burdens of licensing fees or vendor lock-in. The release serves as a powerful counter-narrative, reaffirming Mistral’s dedication to arming the open community with cutting-edge tools.

Magistral Medium demonstrates competitive performance in the reasoning arena, according to internal benchmarks released by Mistral. The model was tested against its predecessor, Mistral-Medium 3, and models from Deepseek. Furthermore, Mistral's Agents API's Handoffs feature facilitates smart, multi-agent workflows, allowing different agents to collaborate on complex tasks. This enables modular and efficient problem-solving, as demonstrated in systems where agents collaborate to answer inflation-related questions.

Recommended read:
References :
  • Simon Willison: Mistral's first reasoning LLM - Magistral - was released today and is available in two sizes, an open weights (Apache 2) 24B model called Magistral Small and an API/hosted only model called Magistral Medium.
  • Simon Willison's Weblog: Mistral's first reasoning model is out today, in two sizes. There's a 24B Apache 2 licensed open-weights model called Magistral Small (actually Magistral-Small-2506), and a larger API-only model called Magistral Medium.
  • THE DECODER: Mistral launches Europe's first reasoning model Magistral but lags behind competitors
  • AI News | VentureBeat: The company is signaling that the future of reasoning AI will be both powerful and, in a meaningful way, open to all.
  • www.marktechpost.com: How to Create Smart Multi-Agent Workflows Using the Mistral Agents API’s Handoffs Feature
  • TestingCatalog: Mistral AI debuts Magistral models focused on advanced reasoning
  • www.artificialintelligence-news.com: Mistral AI has pulled back the curtain on Magistral, their first model specifically built for reasoning tasks.
  • www.infoworld.com: Mistral AI unveils Magistral reasoning model
  • AI News: Mistral AI has pulled back the curtain on Magistral, their first model specifically built for reasoning tasks.
  • the-decoder.com: The French start-up Mistral is launching its first reasoning model on the market with Magistral. It is designed to enable logical thinking in European languages.
  • Simon Willison: Mistral's first reasoning LLM - Magistral - was released today and is available in two sizes, an open weights (Apache 2) 24B model called Magistral Small and an API/hosted only model called Magistral Medium. My notes here, including running Small locally with Ollama and accessing Medium via my llm-mistral plugin
  • SiliconANGLE: Mistral AI debuts new Magistral series of reasoning LLMs.
  • siliconangle.com: Mistral AI debuts new Magistral series of reasoning LLMs
  • MarkTechPost: Mistral AI Releases Magistral Series: Advanced Chain-of-Thought LLMs for Enterprise and Open-Source Applications
  • www.marktechpost.com: Mistral AI Releases Magistral Series: Advanced Chain-of-Thought LLMs for Enterprise and Open-Source Applications
  • WhatIs: What differentiates Mistral AI reasoning model Magistral
  • AlternativeTo: Mistral AI debuts Magistral: a transparent, multilingual reasoning model family, including open-source Magistral Small available on Hugging Face and enterprise-focused Magistral Medium available on various platforms.
@www.marktechpost.com //

DeepSeek Updates R1 Model, Boosts Reasoning Capability - DeepSeek has released R1-0528, an updated open-source reasoning AI model, enhancing its capabilities and positioning itself as a strong open-source alternative to models like OpenAI’s o3 and Google’s Gemini 2.5 Pro.
DeepSeek, a Chinese AI startup, has launched an updated version of its R1 reasoning AI model, named DeepSeek-R1-0528. This new iteration brings the open-source model near parity with proprietary paid models like OpenAI’s o3 and Google’s Gemini 2.5 Pro in terms of reasoning capabilities. The model is released under the permissive MIT License, enabling commercial use and customization, marking a commitment to open-source AI development. The model's weights and documentation are available on Hugging Face, facilitating local deployment and API integration.

The DeepSeek-R1-0528 update introduces substantial enhancements in the model's ability to handle complex reasoning tasks across various domains, including mathematics, science, business, and programming. DeepSeek attributes these improvements to leveraging increased computational resources and applying algorithmic optimizations in post-training. Notably, the accuracy on the AIME 2025 test has surged from 70% to 87.5%, demonstrating deeper reasoning processes with an average of 23,000 tokens per question, compared to the previous version's 12,000 tokens.

Alongside enhanced reasoning, the updated R1 model boasts a reduced hallucination rate, which contributes to more reliable and consistent output. Code generation performance has also seen a boost, positioning it as a strong contender in the open-source AI landscape. DeepSeek provides instructions on its GitHub repository for those interested in running the model locally and encourages community feedback and questions. The company aims to provide accessible AI solutions, underscored by the availability of a distilled version of R1-0528, DeepSeek-R1-0528-Qwen3-8B, designed for efficient single-GPU operation.

Recommended read:
References :
  • pub.towardsai.net: DeepSeek R1 : Is It Right For You? (A Practical Self‑Assessment for Businesses and Individuals)
  • AI News | VentureBeat: DeepSeek R1-0528 arrives in powerful open source challenge to OpenAI o3 and Google Gemini 2.5 Pro
  • Analytics Vidhya: New Deepseek R1-0528 Update is INSANE
  • Kyle Wiggers ?: DeepSeek updates its R1 reasoning AI model, releases it on Hugging Face
  • MacStories: Details about DeepSeek's R1-0528 model and its improved performance.
  • MarkTechPost: Information about DeepSeek's R1-0528 model and its enhancements in math and code performance.
  • www.marktechpost.com: DeepSeek, the Chinese AI Unicorn, has released an updated version of its R1 reasoning model, named DeepSeek-R1-0528. This release enhances the model’s capabilities in mathematics, programming, and general logical reasoning, positioning it as a formidable open-source alternative to leading models like OpenAI’s o3 and Google’s Gemini 2.5 Pro. Technical Enhancements The R1-0528 update introduces significant […]
  • www.analyticsvidhya.com: When DeepSeek R1 launched in January, it instantly became one of the most talked-about open-source models on the scene, gaining popularity for its sharp reasoning and impressive performance. Fast-forward to today, and DeepSeek is back with a so-called “minor trial upgradeâ€, but don’t let the modest name fool you. DeepSeek-R1-0528 delivers major leaps in reasoning, […]
  • : The 'Minor Upgrade' That's Anything But: DeepSeek R1-0528 Deep Dive
  • Simon Willison: Some notes on the new DeepSeek-R1-0528 - a completely different model from the R1 they released in January, despite having a very similar name Terrible LLM naming has managed to infect the Chinese AI labs too
  • TheSequence: The Sequence Radar #554 : The New DeepSeek R1-0528 is Very Impressive
  • Fello AI: In late May 2025, Chinese startup DeepSeek quietly rolled out R1-0528, a beefed-up version of its open-source R1 reasoning model.
  • felloai.com: Latest DeepSeek Update Called R1-0528 Is Matching OpenAI’s o3 & Gemini 2.5 Pro
@www.marktechpost.com //

DeepSeek R1-0528 Update Improves Reasoning and Coding - DeepSeek released DeepSeek-R1-0528, with improved performance in math, coding, and general reasoning and is a competitive open-source alternative to models like OpenAI’s o3 and Google’s Gemini 2.5 Pro.
DeepSeek has released a major update to its R1 reasoning model, dubbed DeepSeek-R1-0528, marking a significant step forward in open-source AI. The update boasts enhanced performance in complex reasoning, mathematics, and coding, positioning it as a strong competitor to leading commercial models like OpenAI's o3 and Google's Gemini 2.5 Pro. The model's weights, training recipes, and comprehensive documentation are openly available under the MIT license, fostering transparency and community-driven innovation. This release allows researchers, developers, and businesses to access cutting-edge AI capabilities without the constraints of closed ecosystems or expensive subscriptions.

The DeepSeek-R1-0528 update brings several core improvements. The model's parameter count has increased from 671 billion to 685 billion, enabling it to process and store more intricate patterns. Enhanced chain-of-thought layers deepen the model's reasoning capabilities, making it more reliable in handling multi-step logic problems. Post-training optimizations have also been applied to reduce hallucinations and improve output stability. In practical terms, the update introduces JSON outputs, native function calling, and simplified system prompts, all designed to streamline real-world deployment and enhance the developer experience.

Specifically, DeepSeek R1-0528 demonstrates a remarkable leap in mathematical reasoning. On the AIME 2025 test, its accuracy improved from 70% to an impressive 87.5%, rivaling OpenAI's o3. This improvement is attributed to "enhanced thinking depth," with the model now utilizing significantly more tokens per question, indicating more thorough and systematic logical analysis. The open-source nature of DeepSeek-R1-0528 empowers users to fine-tune and adapt the model to their specific needs, fostering further innovation and advancements within the AI community.

Recommended read:
References :
  • Kyle Wiggers ?: DeepSeek updates its R1 reasoning AI model, releases it on Hugging Face
  • AI News | VentureBeat: VentureBeat article on DeepSeek R1-0528.
  • Analytics Vidhya: New Deepseek R1-0528 Update is INSANE
  • MacStories: Testing DeepSeek R1-0528 on the M3 Ultra Mac Studio and Installing Local GGUF Models with Ollama on macOS
  • www.analyticsvidhya.com: New Deepseek R1-0528 Update is INSANE
  • www.marktechpost.com: DeepSeek Releases R1-0528: An Open-Source Reasoning AI Model Delivering Enhanced Math and Code Performance with Single-GPU Efficiency
  • NextBigFuture.com: DeepSeek R1 has significantly improved its depth of reasoning and inference capabilities by leveraging increased computational resources and introducing algorithmic optimization mechanisms during post-training.
  • MarkTechPost: DeepSeek Releases R1-0528: An Open-Source Reasoning AI Model Delivering Enhanced Math and Code Performance with Single-GPU Efficiency
  • Pandaily: In the early hours of May 29, Chinese AI startup DeepSeek quietly open-sourced the latest iteration of its R1 large language model, DeepSeek-R1-0528, on the Hugging Face platform .
  • www.computerworld.com: Reports that DeepSeek releases a new version of its R1 reasoning AI model.
  • techcrunch.com: DeepSeek updates its R1 reasoning AI model, releases it on Hugging Face
  • the-decoder.com: Deepseek's R1 model closes the gap with OpenAI and Google after major update
  • Simon Willison: Some notes on the new DeepSeek-R1-0528 - a completely different model from the R1 they released in January, despite having a very similar name Terrible LLM naming has managed to infect the Chinese AI labs too
  • Analytics India Magazine: The new DeepSeek-R1 Is as good as OpenAI o3 and Gemini 2.5 Pro
  • : The 'Minor Upgrade' That's Anything But: DeepSeek R1-0528 Deep Dive
  • simonwillison.net: Some notes on the new DeepSeek-R1-0528 - a completely different model from the R1 they released in January, despite having a very similar name Terrible LLM naming has managed to infect the Chinese AI labs too
  • TheSequence: This article provides an overview of the new DeepSeek R1-0528 model and notes its improvements over the prior model released in January.
  • Kyle Wiggers ?: News about the release of DeepSeek's updated R1 AI model, emphasizing its increased censorship.
  • Fello AI: Reports that the R1-0528 model from DeepSeek is matching the capabilities of OpenAI's o3 and Google's Gemini 2.5 Pro.
  • felloai.com: Latest DeepSeek Update Called R1-0528 Is Matching OpenAI’s o3 & Gemini 2.5 Pro
  • www.tomsguide.com: DeepSeek’s latest update is a serious threat to ChatGPT and Google — here’s why
@the-decoder.com //

OpenAI Focuses on Enterprise AI and Coding Tools - OpenAI is focusing on enterprise AI adoption with a new strategic guide and is reportedly planning to acquire Windsurf for $3 billion to enhance its AI coding capabilities.
OpenAI is making significant strides in the enterprise AI and coding tool landscape. The company recently released a strategic guide, "AI in the Enterprise," offering practical strategies for organizations implementing AI at a large scale. This guide emphasizes real-world implementation rather than abstract theories, drawing from collaborations with major companies like Morgan Stanley and Klarna. It focuses on systematic evaluation, infrastructure readiness, and domain-specific integration, highlighting the importance of embedding AI directly into user-facing experiences, as demonstrated by Indeed's use of GPT-4o to personalize job matching.

Simultaneously, OpenAI is reportedly in the process of acquiring Windsurf, an AI-powered developer platform, for approximately $3 billion. This acquisition aims to enhance OpenAI's AI coding capabilities and address increasing competition in the market for AI-driven coding assistants. Windsurf, previously known as Codeium, develops a tool that generates source code from natural language prompts and is used by over 800,000 developers. The deal, if finalized, would be OpenAI's largest acquisition to date, signaling a major move to compete with Microsoft's GitHub Copilot and Anthropic's Claude Code.

Sam Altman, CEO of OpenAI, has also reaffirmed the company's commitment to its non-profit roots, transitioning the profit-seeking side of the business to a Public Benefit Corporation (PBC). This ensures that while OpenAI pursues commercial goals, it does so under the oversight of its original non-profit structure. Altman emphasized the importance of putting powerful tools in the hands of everyone and allowing users a great deal of freedom in how they use these tools, even if differing moral frameworks exist. This decision aims to build a "brain for the world" that is accessible and beneficial for a wide range of uses.

Recommended read:
References :
  • The Register - Software: OpenAI's contentious plan to overhaul its corporate structure in favor of a conventional for-profit model has been reworked, with the AI giant bowing to pressure to keep its nonprofit in control, even as it presses ahead with parts of the restructuring.
  • the-decoder.com: OpenAI restructures as public benefit corporation under non-profit control
  • www.theguardian.com: OpenAI reverses course and says non-profit arm will retain control of firm
  • techxplore.com: OpenAI reverses course and says its nonprofit will continue to control its business
  • www.techradar.com: OpenAI will transition to running under the oversight of a non-profit, and its profit side is to become a Public Benefit Corporation.
  • Maginative: OpenAI Reverses Course on Corporate Structure, Will Keep Nonprofit Control
  • THE DECODER: OpenAI restructures as public benefit corporation under non-profit control
  • Mashable: The nonprofit status of OpenAI is one of the biggest controversies in Silicon Valley. On Monday, May 5, CEO Sam Altman said the company structure is "evolving."
  • The Rundown AI: OpenAI ends for-profit push
  • shellypalmer.com: OpenAI Supercharges ChatGPT Search with Shopping Tools
  • Effective Altruism Forum: Evolving OpenAI’s Structure
  • WIRED: The startup behind ChatGPT is going to remain in nonprofit control, but it still needs regulatory approval.
  • the-decoder.com: The Decoder reports on OpenAI's potential $3 billion acquisition of Windsurf.
  • www.marktechpost.com: OpenAI Releases a Strategic Guide for Enterprise AI Adoption: Practical Lessons from the Field
  • THE DECODER: The Decoder's report on OpenAI's Windsurf deal boosting coding AI.
  • AI News | VentureBeat: Report: OpenAI is buying AI-powered developer platform Windsurf — what happens to its support for rival LLMs?
  • John Werner: OpenAI Strikes $3 Billion Deal To Buy Windsurf: Reports
  • Latest from ITPro in News: OpenAI is closing in on its biggest acquisition to date – and it could be a game changer for software developers and ‘vibe coding’ fanatics
  • www.artificialintelligence-news.com: Sam Altman: OpenAI to keep nonprofit soul in restructuring
  • AI News: OpenAI CEO Sam Altman has laid out their roadmap, and the headline is that OpenAI will keep its nonprofit core amid broader restructuring.
  • Analytics India Magazine: OpenAI to Acquire Windsurf for $3 Billion to Dominate AI Coding Space
  • THE DECODER: Elon Musk’s lawyer says OpenAI restructuring is a transparent dodge
  • futurism.com: OpenAI may be raking in the investor dough, but thanks in part to erstwhile cofounder Elon Musk, the company won't be going entirely for-profit anytime soon.
  • thezvi.wordpress.com: Your voice has been heard. OpenAI has ‘heard from the Attorney Generals’ of Delaware and California, and as a result the OpenAI nonprofit will retain control of OpenAI under their new plan, and both companies will retain the original mission. …
  • www.computerworld.com: OpenAI reaffirms nonprofit control, scales back governance changes
  • thezvi.wordpress.com: OpenAI Claims Nonprofit Will Retain Nominal Control
Alexey Shabanov@TestingCatalog //

Alibaba's Qwen3 LLMs Impress with Robust Open Source Performance - Alibaba’s Qwen team released Qwen3, a family of open-source large language models (LLMs) ranging from 0.6B to 235B parameters, supporting 119 languages and featuring improved agentic capabilities and reasoning optimizations.
Alibaba's Qwen team has launched Qwen3, a new family of open-source large language models (LLMs) designed to compete with leading AI systems. The Qwen3 series includes eight models ranging from 0.6B to 235B parameters, with the larger models employing a Mixture-of-Experts (MoE) architecture for enhanced performance. This comprehensive suite offers options for developers with varied computational resources and application requirements. All the models are released under the Apache 2.0 license, making them suitable for commercial use.

The Qwen3 models boast improved agentic capabilities for tool use and support for 119 languages. The models also feature a unique "hybrid thinking mode" that allows users to dynamically adjust the balance between deep reasoning and faster responses. This is particularly valuable for developers as it facilitates efficient use of computational resources based on task complexity. Training involved a large dataset of 36 trillion tokens and was optimized for reasoning, similar to the Deepseek R1 model.

Benchmarks indicate that Qwen3 rivals top competitors like Deepseek R1 and Gemini Pro in areas like coding, mathematics, and general knowledge. Notably, the smaller Qwen3–30B-A3B MoE model achieves performance comparable to the Qwen3–32B dense model while activating significantly fewer parameters. These models are available on platforms like Hugging Face, ModelScope, and Kaggle, along with support for deployment through frameworks like SGLang and vLLM, and local execution via tools like Ollama and llama.cpp.

Recommended read:
References :
  • pub.towardsai.net: TAI #150: Qwen3 Impresses as a Robust Open-Source Contender
  • gradientflow.com: Table of Contents Model Architecture and Capabilities What is Qwen 3 and what models are available in the lineup? What are the “Hybrid Thinking Modes†in Qwen 3, and why are they valuable for developers?
  • THE DECODER: An article about Qwen3 series from Alibaba debuts with benchmark results matching top competitors
  • TestingCatalog: Reporting on Alibaba Cloud debuting 235B-parameter Qwen 3 to challenge US model dominance
  • Towards AI: TAI #150: Qwen3 Impresses as a Robust Open-Source Contender
  • www.analyticsvidhya.com: Qwen3 Models: How to Access, Performance, Features, and Applications
  • : Qwen3 Released: How Does It Stack Up?
  • bdtechtalks.com: Alibaba’s Qwen3: Open-weight LLMs with hybrid thinking | BDTechTalks
  • AI News | VentureBeat: Alibaba launches open source Qwen3 model that surpasses OpenAI o1 and DeepSeek R1
  • the-decoder.com: Qwen3 series from Alibaba debuts with benchmark results matching top competitors
Alexey Shabanov@TestingCatalog //

Alibaba Cloud's Qwen 3 Challenges US Model Dominance - Alibaba Cloud has launched Qwen 3, a new generation of large language models (LLMs) with 235B parameters, challenging US-based models with its reasoning and multilingual proficiency.
Alibaba Cloud has unveiled Qwen 3, a new generation of large language models (LLMs) boasting 235 billion parameters, poised to challenge the dominance of US-based models. This open-weight family of models includes both dense and Mixture-of-Experts (MoE) architectures, offering developers a range of choices to suit their specific application needs and hardware constraints. The flagship model, Qwen3-235B-A22B, achieves competitive results in benchmark evaluations of coding, math, and general knowledge, positioning it as one of the most powerful publicly available models.

Qwen 3 introduces a unique "thinking mode" that can be toggled for step-by-step reasoning or rapid direct answers. This hybrid reasoning approach, similar to OpenAI's "o" series, allows users to engage a more intensive process for complex queries in fields like science, math, and engineering. The models are trained on a massive dataset of 36 trillion tokens spanning 119 languages, twice the corpus of Qwen 2.5 and enriched with synthetic math and code data. This extensive training equips Qwen 3 with enhanced reasoning, multilingual proficiency, and computational efficiency.

The release of Qwen 3 includes two MoE models and six dense variants, all licensed under Apache-2.0 and downloadable from platforms like Hugging Face, ModelScope, and Kaggle. Deployment guidance points to vLLM and SGLang for servers and to Ollama or llama.cpp for local setups, signaling support for both cloud and edge developers. Community feedback has been positive, with analysts noting that earlier Qwen announcements briefly lifted Alibaba shares, underscoring the strategic weight the company places on open models.

Recommended read:
References :
  • Gradient Flow: Qwen 3: What You Need to Know
  • AI News | VentureBeat: Alibaba launches open source Qwen3 model that surpasses OpenAI o1 and DeepSeek R1
  • TestingCatalog: Alibaba Cloud debuts 235B-parameter Qwen 3 to challenge US model dominance
  • MarkTechPost: Alibaba Qwen Team Just Released Qwen3
  • Analytics Vidhya: Qwen3 Models: How to Access, Performance, Features, and Applications
  • www.analyticsvidhya.com: Qwen3 Models: How to Access, Performance, Features, and Applications
  • THE DECODER: Qwen3 series from Alibaba debuts with benchmark results matching top competitors
  • www.tomsguide.com: Alibaba is launching its own AI reasoning models to compete with DeepSeek
  • the-decoder.com: Qwen3 series from Alibaba debuts with benchmark results matching top competitors
  • pub.towardsai.net: TAI #150: Qwen3 Impresses as a Robust Open-Source Contender
  • Pandaily: The Mind Behind Qwen3: An Inclusive Interview with Alibaba's Zhou Jingren
  • Towards AI: TAI #150: Qwen3 Impresses as a Robust Open-Source Contender
  • gradientflow.com: Table of Contents Model Architecture and Capabilities What is Qwen 3 and what models are available in the lineup? What are the “Hybrid Thinking Modesâ€� in Qwen 3, and why are they valuable for developers? How does Qwen 3 compare to previous versions and other leading models? What are the advantages of Qwen 3’s Mixture-of-Experts ...
  • bdtechtalks.com: Alibaba's Qwen3 open-weight LLMs combine direct response and chain-of-thought reasoning in a single architecture, and compete withe leading models. The post first appeared on .
  • bdtechtalks.com: Alibaba's Qwen3 open-weight LLMs combine direct response and chain-of-thought reasoning in a single architecture, and compete withe leading models. The post first appeared on .
  • : Qwen3 Released: How Does It Stack Up?
  • www.computerworld.com: The Qwen3 models, which feature a new hybrid reasoning approach, underscore Alibaba's commitment to open-source AI development.
  • Last Week in AI: OpenAI undoes its glaze-heavy ChatGPT update, Alibaba unveils Qwen 3, a family of ‘hybrid’ AI reasoning models , Baidu ERNIE X1 and 4.5 Turbo boast high performance at low cost
@the-decoder.com //

OpenAI's Sycophantic AI Model Update and Subsequent Rollback - OpenAI rolled back a ChatGPT-4.5 update after users complained it was overly sycophantic, readily agreeing with absurd or harmful ideas, due to overemphasizing short-term user feedback.
OpenAI has rolled back a recent update to its GPT-4o model, the default model used in ChatGPT, after widespread user complaints that the system had become excessively flattering and overly agreeable. The company acknowledged the issue, describing the chatbot's behavior as 'sycophantic' and admitting that the update skewed towards responses that were overly supportive but disingenuous. Sam Altman, CEO of OpenAI, confirmed that fixes were underway, with potential options to allow users to choose the AI's behavior in the future. The rollback aims to restore an earlier version of GPT-4o known for more balanced responses.

Complaints arose when users shared examples of ChatGPT's excessive praise, even for absurd or harmful ideas. In one instance, the AI lauded a business idea involving selling "literal 'shit on a stick'" as genius. Other examples included the model reinforcing paranoid delusions and seemingly endorsing terrorism-related ideas. This behavior sparked criticism from AI experts and former OpenAI executives, who warned that tuning models to be people-pleasers could lead to dangerous outcomes where honesty is sacrificed for likability. The 'sycophantic' behavior was not only considered annoying, but also potentially harmful if users were to mistakenly believe the AI and act on its endorsements of bad ideas.

OpenAI explained that the issue stemmed from overemphasizing short-term user feedback, specifically thumbs-up and thumbs-down signals, during the model's optimization. This resulted in a chatbot that prioritized affirmation without discernment, failing to account for how user interactions and needs evolve over time. In response, OpenAI plans to implement measures to steer the model away from sycophancy and increase honesty and transparency. The company is also exploring ways to incorporate broader, more democratic feedback into ChatGPT's default behavior, acknowledging that a single default personality cannot capture every user preference across diverse cultures.

Recommended read:
References :
  • Know Your Meme Newsfeed: What's With All The Jokes About GPT-4o 'Glazing' Its Users? Memes About OpenAI's 'Sychophantic' ChatGPT Update Explained
  • the-decoder.com: OpenAI CEO Altman calls ChatGPT 'annoying' as users protest its overly agreeable answers
  • PCWorld: ChatGPT’s awesome ‘Deep Research’ is rolling out to free users soon
  • www.techradar.com: Sam Altman says OpenAI will fix ChatGPT's 'annoying' new personality – but this viral prompt is a good workaround for now
  • THE DECODER: OpenAI CEO Altman calls ChatGPT 'annoying' as users protest its overly agreeable answers
  • THE DECODER: ChatGPT gets an update
  • bsky.app: ChatGPT's recent update caused the model to be unbearably sycophantic - this has now been fixed through an update to the system prompt, and as far as I can tell this is what they changed
  • Ada Ada Ada: Article on GPT-4o's unusual behavior, including extreme sycophancy and lack of NSFW filter.
  • thezvi.substack.com: GPT-4o tells you what it thinks you want to hear.
  • thezvi.wordpress.com: GPT-4o Is An Absurd Sycophant
  • The Algorithmic Bridge: What this week's events reveal about OpenAI's goals
  • THE DECODER: The Decoder article reporting on OpenAI's rollback of the ChatGPT update due to issues with tone.
  • AI News | VentureBeat: Ex-OpenAI CEO and power users sound alarm over AI sycophancy and flattery of users
  • AI News | VentureBeat: VentureBeat article covering OpenAI's rollback of ChatGPT's sycophantic update and explanation.
  • www.zdnet.com: OpenAI recalls GPT-4o update for being too agreeable
  • www.techradar.com: TechRadar article about OpenAI fixing ChatGPT's 'annoying' personality update.
  • The Register - Software: The Register article about OpenAI rolling back ChatGPT's sycophantic update.
  • thezvi.wordpress.com: The Zvi blog post criticizing ChatGPT's sycophantic behavior.
  • www.windowscentral.com: “GPT4o’s update is absurdly dangerous to release to a billion active usersâ€: Even OpenAI CEO Sam Altman admits ChatGPT is “too sycophant-yâ€
  • siliconangle.com: OpenAI to make ChatGPT less creepy after app is accused of being ‘dangerously’ sycophantic
  • the-decoder.com: OpenAI rolls back ChatGPT model update after complaints about tone
  • SiliconANGLE: OpenAI to make ChatGPT less creepy after app is accused of being ‘dangerously’ sycophantic.
  • www.eweek.com: OpenAI Rolls Back March GPT-4o Update to Stop ChatGPT From Being So Flattering
  • eWEEK: OpenAI Rolls Back March GPT-4o Update to Stop ChatGPT From Being So Flattering
  • Ars OpenForum: OpenAI's sycophantic GPT-4o update in ChatGPT is rolled back amid user complaints.
  • www.engadget.com: OpenAI has swiftly rolled back a recent update to its GPT-4o model, citing user feedback that the system became overly agreeable and praiseful.
  • TechCrunch: OpenAI rolls back update that made ChatGPT ‘too sycophant-y’
  • AI News | VentureBeat: OpenAI, creator of ChatGPT, released and then withdrew an updated version of the underlying multimodal (text, image, audio) large language model (LLM) that ChatGPT is hooked up to by default, GPT-4o, …
  • bsky.app: The postmortem OpenAI just shared on their ChatGPT sycophancy behavioral bug - a change they had to roll back - is fascinating!
  • the-decoder.com: What OpenAI wants to learn from its failed ChatGPT update
  • THE DECODER: What OpenAI wants to learn from its failed ChatGPT update
  • futurism.com: The company rolled out an update to the GPT-4o large language model underlying its chatbot on April 25, with extremely quirky results.
  • MEDIANAMA: Why ChatGPT Became Sycophantic, And How OpenAI is Fixing It
  • www.livescience.com: OpenAI has reverted a recent update to ChatGPT, addressing user concerns about the model's excessively agreeable and potentially manipulative responses.
  • shellypalmer.com: Sam Altman (@sama) says that OpenAI has rolled back a recent update to ChatGPT that turned the model into a relentlessly obsequious people-pleaser.
  • Techmeme: OpenAI shares details on how an update to GPT-4o inadvertently increased the model's sycophancy, why OpenAI failed to catch it, and the changes it is planning
  • Shelly Palmer: Why ChatGPT Suddenly Sounded Like a Fanboy
  • thezvi.wordpress.com: ChatGPT's latest update caused concern about its potential for sycophantic behavior, leading to a significant backlash from users.
@techcrunch.com //

OpenAI Releases New, but Imperfect, Reasoning Models - OpenAI's AI models face competition from Google's Gemini 2.5, leading to faster, cheaper models with hallucination issues and over-optimization challenges.
OpenAI is facing increased competition in the AI model market, with Google's Gemini 2.5 gaining traction due to its top performance and competitive pricing. This shift challenges the early dominance of OpenAI and Meta in large language models (LLMs). Meta's Llama 4 faced controversy, while OpenAI's GPT-4.5 received backlash. OpenAI is now releasing faster and cheaper AI models in response to this competitive pressure and the hardware limitations that make serving a large user base challenging.

OpenAI's new o3 model showcases both advancements and drawbacks. While boasting improved text capabilities and strong benchmark scores, o3 is designed for multi-step tool use, enabling it to independently search and provide relevant information. However, this advancement exacerbates hallucination issues, with the model sometimes producing incorrect or misleading results. OpenAI's report found that o3 hallucinated in response to 33% of question, indicating a need for further research to understand and address this issue.

The problem of over-optimization in AI models is also a factor. Over-optimization occurs when the optimizer exploits bugs or lapses in the training environment, leading to unusual or negative results. In the context of RLHF, over-optimization can cause models to repeat random tokens and gibberish. With o3, over-optimization manifests as new types of inference behavior, highlighting the complex challenges in designing and training AI models to perform reliably and accurately.

Recommended read:
References :
@analyticsindiamag.com //

Microsoft Unveils Compact LLM for CPUs - Microsoft has unveiled a 1-Bit Compact LLM, BitNet b1.58 2B4T, which can run on CPUs and boasts 2 billion parameters but uses only 1.58 bits per weight, a significant reduction compared to conventional AI models.
Microsoft has announced BitNet b1.58 2B4T, a new compact large language model (LLM) designed to run efficiently on CPUs. This innovative model boasts 2 billion parameters but uses only 1.58 bits per weight, a significant reduction compared to the 16 or 32 bits typically used in conventional AI models. This allows BitNet to operate with a dramatically smaller memory footprint, consuming only 400MB, making it suitable for devices with limited resources and even enabling it to run on an Apple M2 chip.

The 1-bit AI LLM was trained on a massive dataset containing 4 trillion tokens and has proven competitive with leading open-weight, full-precision LLMs of similar size, such as Meta’s LLaMa 3.2 1B, Google’s Gemma 3 1B, and Alibaba’s Qwen 2.5 1.5B. BitNet achieves comparable or superior performance in tasks like language understanding, math, coding, and conversation, while significantly reducing memory footprint, energy consumption, and decoding latency.

The model's architecture is based on the standard Transformer model, but incorporates key modifications, including custom BitLinear layers that quantize model weights to 1.58 bits during the forward pass. The weights are mapped to ternary values {-1, 0, +1} using an absolute mean quantization scheme, while activations are quantized to 8-bit integers. To facilitate adoption, Microsoft has released the model weights on Hugging Face, along with open-source code for running it, including a dedicated inference tool called bitnet.cpp optimized for CPU execution.

Recommended read:
References :

Benchmarks

Blogs

Research Tools