arXiv:2606.30815v1 Announce Type: new Abstract: Recent work suggests that transformer language models show a bias towards human languages over unnatural ("impossible") languages argued to be unacquirable by humans. However, this literature has largely based these claims on differences in sample efficiency and test-set perplexity, rather than on direct evaluations of the linguistic capacities that could plausibly explain non-attestation in human languages. We evaluate two theoretically motivated linking hypotheses: impossibility arising from deficiencies in grammatical sensitivity or generative production. Using GPT-2 style models trained on perturbed "impossible" variants of English, we measure sensitivity to grammaticality using BLiMP minimal pairs, finding that model performance exhibits only gradual degradation, mediated by the language's information locality. In contrast, these models exhibited pronounced failures in generation, producing substantially fewer high-quality sentences at longer lengths. Together, these results suggest generative deficiency and transmission failures as a plausible linking hypothesis between language model behaviour and non-attestation of impossible languages.

arXiv:2606.30851v1 Announce Type: new Abstract: Improving the reliability of large language models (LLMs) at inference time is a central challenge in structured reasoning tasks such as Text-to-SQL. Common test-time inference strategies, including Best-of-N sampling and Majority Voting, rely on heuristic signals such as execution success or output frequency, which provide limited semantic discrimination across candidate outputs. In this work, we study Outcome Reward Models (ORMs) as learned semantic scoring functions for test-time verification in Text-to-SQL. While ORMs have been previously explored for test-time scaling and alignment, their application to structured query generation remains underexplored. We introduce GradeSQL, a scalable framework for training task-specific ORMs via automated candidate generation and execution-based labeling, enabling verifier training without manual annotation. We integrate ORMs into a verification-driven Best-of-N pipeline and evaluate our approach on the BIRD and Spider benchmarks across multiple open-source LLM families. ORM-based selection consistently outperforms execution-based Best-of-N and Majority Voting, with gains of up to +4.33% on BIRD and +2.10% on Spider. We further show that ORMs scale effectively with larger candidate sets and yield stronger improvements on complex queries. Overall, our results demonstrate that ORM-based verification provides a simple, effective, and scalable alternative to heuristic test-time selection strategies for Text-to-SQL. Code datasets and models are publicly available.

arXiv:2606.28386v1 Announce Type: new Abstract: Image autoregressive models (IARs) have recently demonstrated remarkable capabilities in visual content generation, achieving photorealistic quality and rapid synthesis through the next-token prediction paradigm adapted from large language models. As these models become widely accessible, robust data provenance is required to reliably trace IAR-generated images to the source model that synthesized them. This is critical to prevent the spread of misinformation, detect fraud, and attribute harmful content. We find that although IAR-generated images often appear visually identical to real images, their generation process introduces characteristic patterns in their outputs, which serves as a reliable provenance signal for the generated images. Leveraging this, we present a post-hoc framework that enables the robust detection of such patterns for provenance tracing. Notably, our framework does not require modifications of the generative process or outputs. Thereby, it is applicable in contexts where prior watermarking methods cannot be used, such as for generated content that is already published without additional marks and for models that do not integrate watermarking. We demonstrate the effectiveness of our approach across a wide range of IARs, highlighting its high potential for robust data provenance tracing in autoregressive image generation.

arXiv:2606.30705v1 Announce Type: new Abstract: Deterministic few-step generation succeeds on continuous image latents but collapses to incoherent text on continuous text latents, and we show the cause is geometric rather than a training or scaling deficiency: a smooth, regularity-limited deterministic map cannot resolve a discrete branch choice before a sharp categorical readout, so few-step failure is governed by decoder sharpness, not transport accuracy. In the overlapping regime of real text autoencoders, we prove (Theorem 3) that the posterior-mean terminal step flips tokens at the rate of the latent mass in an $O(s(t))$ tube around decision boundaries. Two diagnostics, DABI (readout sharpness) and CCI (categorical commitment), measured on published checkpoints show that four independently built continuous-text decoders amplify a boundary-aligned perturbation far beyond a norm-matched isotropic one (DABI from $5\times10^{2}$ to $>10^{5}$), while image decoders have DABI $\approx 1$. Two mechanisms escape the continuous bound: categorical commitment (autoregressive decoders succeed despite sharper readouts) and stochastic re-injection (deterministic ODE at $K=4$ gives PPL 294 versus SDE 50 on the same model). In the idealized separated regime we prove matching sharp transport laws, including a dimension phase diagram: the deterministic stiffness needed to separate $M$ modes grows as $\Theta(\sqrt{\log M})$ once the latent dimension is $\Omega(\log M)$ (and as $M^{1/n}$ in fixed dimension), with a depth-$B$ hierarchy giving a $\sqrt{B}$-smaller per-step peak (Theorems 5-7); a coarea identity links these to the overlapping tube (Theorem 17). The result is an accuracy-depth-stiffness tradeoff: within the deterministic-continuous class the cost is irreducible, and both escapes step outside it.

Music from Kylie Minogue, John Farnham, INXS, Midnight Oil, AC/DC, Tones and I, Gotye, Ben Frost, Nick Cave, Tame Impala, Parkway Drive, The Living End and Vance Joy has been found in a database of 12 million songs used to teach artificial intelligence. This database, listing songs available on YouTube, is used by AI systems to train the ability to recognise and create music. AI relies entirely on these massive databases, trained on almost everything ever placed on the internet. And Australia’s inclusion in these databases is huge. Kylie alone has 182 songs in just one database. The volume of Australian music used to train AI has caused significant anger in the Australian music industry, driven by the knowledge that “AI for music creators” platforms such as Suno create as much music as Spotify’s entire catalogue every two weeks. Dobe Newton, co-writer of folk classic I am Australian and member of the Bushwackers, has music included in the databases. He believes there is “no real ethical nor moral underpinning” to current AI music practices. Jesse Pattinson of The Delta Riggs is concerned about “the opportunity it will take away from real artists”. Screen composer and APRA board member Caitlin Yeo told me she holds “deep concern for the future of music made by humans for humans”. She described feeling “violated” when discovering her work in these databases, realising decades of her work had been “hoovered up in a second” to “feed companies offshore that pay no taxes”. While Australian artists are feeling ripped off, the intersection between copyright law and AI makes proving infringements incredibly difficult. Copyright and AI Lawsuits across multiple creative industries have covered how copyrighted books, speeches and even pornography have been used to train AI models. These actions have led to massive settlements. In 2025, AI company Anthropic paid US$1.5 billion to writers who brought a class action lawsuit over the company’s use of a database of more than 7 million books to train its AI. But when it comes to proving copyright infringement, the devil is very much in the details, and it may be harder for the impacted musicians to prove infringement than authors. Crucially, these databases often do not contain any copyrighted material. Rather, they contain instructions on where to download the data from, along with associated information to help AI training. Legal challenges The distinction about how the databases are packaged up is not a minor thing. Previous lawsuits surrounding similar databases found simply listing where copyrighted material can be found is not, by itself, copyright infringement. The infringement only takes place when an AI company uses the data to train its model. Think of these databases like the map to a safe filled with gold. Having the map itself is fine; stealing the gold is when the law is broken. While we know at least some AI companies have used this data map, and that their training relies upon it, legally establishing use and any copyright infringements is a challenge. This legal battlefield is complicated by the unique nature of music law. Copyright protects specific expressions, like a distinct melody or recording, but not a general style. AI developers exploit this loophole. Rather than copying note-for-note, they extract underlying patterns, chord progressions and vocal textures to create a pastiche. To a creator, this feels like theft. But in the eyes of the law it may just be imitation. Legal challenges aren’t impossible. The German music royalties society successfully sued chatGPT’s OpenAI. Universal and Sony, representing more than 50% of the music industry, have sued Suno for infringing the copyright of more than 60,000 songs, following other cases against AI giant NVIDIA and Suno. In response to these lawsuits, Suno described its platform as a fair-use training model, saying “learning is not infringing”. What’s next? Since artists became aware of these datasets, the international music industry has rapidly coalesced around several class action lawsuits and lobbying efforts. In Australia, 4,000 artists signed a petition calling on the government to increase protections for artists and today, artists held a press conference in Canberra to discuss AI’s impact. Tech lobbyists are arguing for exemptions from Australian copyright law. Similarly the Australian Strategic Policy Institute argues that copyright law is a “strategic liability” that increases Australia’s reliance on foreign AI models. There is a fierce tension between fostering innovation and protecting creative industries. But Australia does not have to choose between the two. The European Union’s AI act forged an alternative path, to prevent AI companies from hiding copyright infringements from artists and rights holders. From next August, all AI models accessed from within the EU must declare the source of their training data, and comply with local copyright laws, no matter where the AI was built. These laws may break the veil of secrecy surrounding the AI data usage, and will significantly increase the likelihood of artists being paid when their work is included in these AI models. The cultural heartbeat of Australia depends upon supporting creatives. Joining or drawing inspiration from the EU’s AI legislation could help protect artists, and ensure they are fairly represented regarding AI works derived from their labour. At the very least, as Yeo told me, artists “should see a slice of the pie too”. Andrew Cullen is a board member of Music Technology Australia and is a member of the World Intellectual Property Organisation's Artificial Intelligence Infrastructure Interchange.

Vinton Cerf, one of the creators of the protocols underlying the internet, will step down as Google's chief internet evangelist next week.

The Atlantic: FOIA docs: White House officials used auto-deleting Signal chats even after Trump advised against its use in April 2025, raising federal recordkeeping concerns  —  New records reveal that officials kept using the app, even after the president suggested they stop.

Carmen Arroyo / Bloomberg: SpaceX cuts monthly Starlink prices in half in the Memphis area, as it endures blowback and legal challenges from opponents of its Colossus data centers  —  SpaceX is offering discounts for Starlink internet plans in Memphis, Tennessee, as the Elon Musk-led company endures blowback …

The Trump administration's erratic approach to AI policymaking has left companies across the industry with little clarity about what will govern future model releases.

Wayve’s offering is part of a growing trend of AI startups using employee tenders as a strategic tool to attract and retain talent.

Bloomberg: Lime raised $174M in its US IPO, selling 6.68M shares at $25 each, at the midpoint of the marketed range, giving Lime a market value of $1.6B  —  Neutron Holdings Inc., better known as Lime, and some of its shareholders raised $174 million in a US initial public offering that priced at the midpoint of the marketed range.

Subrat Patnaik / Bloomberg: Vimeo owner Bending Spoons raised $1.68B by selling 58M shares at $29 each, valuing it at ~$18.4B, in one of the largest US IPOs by a European company in 2026  —  Bending Spoons SpA, which acquires struggling software businesses, and some of its backers raised $1.68 billion in a US initial public offering priced above a marketed range.

Lisa Bonos / Inc: Pie, a startup that uses AI to optimize marketing for brick-and-mortar small businesses, emerges from stealth with a $19.5M Series A led by Lightspeed  —  Pie, which spent the past year building quietly, is betting small businesses need help showing up in AI search the way they once needed help showing up on Google.

The group compromised at least 10 regional organizations, including two state-owned entities, and deployed a new backdoor.

Generative artificial intelligence (AI) is changing the fitness industry: people can now ask chatbots to write marathon plans, build gym programs and even adjust workouts based on sleep or heart rate data. For many, AI feels like the future of fitness coaching because it is fast, cheap and readily available. But while AI can be helpful, research suggests it still has limitations, especially when compared with experienced human coaches. So, let’s look at how it all works and the pros and cons. Why are people using AI for training? There is very little research examining exactly why people use AI for exercise programs, but researchers have offered some potential explanations. Firstly, accessibility and cost: a chatbot can create a strength or running program in seconds without you having to wait for an appointment with an exercise professional. Not to mention these can be generated for free. Secondly, availability. There is some research indicating people appreciate rapid feedback in real-time from AI tools. For example, you could ask an AI tool how to change an exercise due to knee pain and get a response in seconds. However, if you are following a program prescribed by a human coach, you may need to wait a day or two before discussing the issue and receiving feedback. What are the benefits and risks? There is a growing body of research looking at the suitability of AI-generated exercise programs across a host of contexts. One study had ChatGPT design an individualised exercise program for five made-up people, which were then evaluated by a team of experts. They concluded the AI tool could provide safe, basic exercise recommendations, but may not provide enough adaptability to ensure long-term progress. Similarly, another study had expert running coaches assess AI-generated running programs. They thought the exercise programs were suitable for novices but not great for trained athletes. The effectiveness of these programs appear to be highly dependent on the level of information provided. In short, the more context you can provide regarding your current capabilities, goals and fitness level, the better the exercise program will be. However, providing such detailed prompts requires a degree of content-specific knowledge that many people don’t have. This may make AI tools less useful to the average person. Finally, it is not clear whether AI systems can fully account for injuries or medical conditions. Health screening is important to keep people safe before exercising and something all exercise professionals should do before writing you a program. If this is being missed, there is the potential for an AI-generated exercise program to be unsafe for your current level of health. Are human trainers better? There is a small body of research comparing AI-generated exercise programs to human generated programs and the results are interesting. One recent study randomly allocated people to one of two groups: a 12-week weight training program under the guidance of ChatGPT or a 12-week program under the guidance of a personal trainer. There were larger increases in muscle size and strength in the personal trainer group. Another compared a five-week AI generated fitness program to a five-week human-generated program. It found the human-generated program led to slightly greater increases in fitness and endurance than the AI program. Finally, a third study compared a ten-week AI generated athletic performance program against a ten-week human generated program on measures of jump performance in volleyball athletes. They found the human program led to slightly greater improvements in jump distance but the same improvements in jump height. Collectively, these studies suggest that while AI-generated exercise programs can improve your fitness, they might be slightly less effective than programs created by human experts. This may be due to their inability to provide real-time feedback and motivation. However, it is also important to note these studies were all published in relatively low-quality journals and had some limitations. So, their findings should be interpreted with caution. AI at the World Cup: smarter tactics, healthy players, safer crowds – but new risks What should you watch out for? If you choose to use AI, there are some key things to keep in mind: treat AI-generated programs as a starting point. Use them to organise your training, but keep in mind you might need to modify the plan if it feels unrealistic or inappropriate avoid increasing training volume or intensity too quickly. Sudden jumps in running distance or lifting intensity can increase injury risk, and this may not be factored into AI generated programs if you are completely new to a gym environment, you may want to spend a couple of sessions with a human trainer to familiarise yourself with good technique before starting your AI-generated program if you are looking to achieve high levels of performance, you might need to consider a human coach to maximise your progress be extra cautious if you have injuries, a chronic disease, or complex goals. Current AI tools may not be able to personalise your program perfectly and it might be safest to see a professional. Hunter Bennett does not work for, consult, own shares in or receive funding from any company or organisation that would benefit from this article, and has disclosed no relevant affiliations beyond their academic appointment.

@anthropicai: Anthropic says the Department of Commerce has lifted export controls on Claude Fable 5 and Mythos 5 and that it will begin restoring access Wednesday  —  We've received notice that the Department of Commerce has lifted export controls on Claude Fable 5 and Mythos 5. We'll begin restoring access tomorrow, and will share an update soon. We're grateful to our users for their patience, and to everyone who worked with us on

Formula aggregations in rollups, run-end encoding for Barrage, and a broad set of engine and client fixes

<blockquote cite="https://twitter.com/anthropicai/status/2072106151890809341"><p>We’ve received notice that the Department of Commerce has lifted export controls on Claude Fable 5 and Mythos 5.</p> <p>We'll begin restoring access tomorrow, and will share an update soon.</p></blockquote> <p class="cite">&mdash; <a href="https://twitter.com/anthropicai/status/2072106151890809341">Anthropic</a>, on Twitter</p> <p>Tags: <a href="https://simonwillison.net/tags/anthropic">anthropic</a>, <a href="https://simonwillison.net/tags/claude">claude</a>, <a href="https://simonwillison.net/tags/generative-ai">generative-ai</a>, <a href="https://simonwillison.net/tags/claude-mythos-fable">claude-mythos-fable</a>, <a href="https://simonwillison.net/tags/ai">ai</a>, <a href="https://simonwillison.net/tags/llms">llms</a></p>

Ashley Stewart / Business Insider: Sources: Microsoft plans to announce job cuts next week affecting less than 2.5% of its 220,000-person workforce, including roles in sales, consulting, and Xbox  —  Microsoft is planning to announce job cuts soon as the tech giant continues efforts to control costs, according to people familiar with the situation.