arXiv:2604.11835v1 Announce Type: new Abstract: Machine learning for tabular data remains constrained by poor schema generalization, a challenge rooted in the lack of semantic understanding of structured variables. This challenge is particularly acute in domains like clinical medicine, where electronic health record (EHR) schemas vary significantly. To solve this problem, we propose Schema-Adaptive Tabular Representation Learning, a novel method that leverages large language models (LLMs) to create transferable tabular embeddings. By transforming structured variables into semantic natural language statements and encoding them with a pretrained LLM, our approach enables zero-shot alignment across unseen schemas without manual feature engineering or retraining. We integrate our encoder into a multimodal framework for dementia diagnosis, combining tabular and MRI data. Experiments on NACC and ADNI datasets demonstrate state-of-the-art performance and successful zero-shot transfer to unseen schemas, significantly outperforming clinical baselines, including board-certified neurologists, in retrospective diagnostic tasks. These results validate our LLM-driven approach as a scalable, robust solution for heterogeneous real-world data, offering a pathway to extend LLM-based reasoning to structured domains.

arXiv:2604.11833v1 Announce Type: new Abstract: Despite the popularity of Convolutional Neural Networks (CNN), the problem of uncertainty quantification (UQ) of CNN has been largely overlooked. Lack of efficient UQ tools severely limits the application of CNN in certain areas, such as medicine, where prediction uncertainty is critically important. Among the few existing UQ approaches that have been proposed for deep learning, none of them has theoretical consistency that can guarantee the uncertainty quality. To address this issue, we propose a novel bootstrap based framework for the estimation of prediction uncertainty. The inference procedure we use relies on convexified neural networks to establish the theoretical consistency of bootstrap. Our approach has a significantly less computational load than its competitors, as it relies on warm-starts at each bootstrap that avoids refitting the model from scratch. We further explore a novel transfer learning method so our framework can work on arbitrary neural networks. We experimentally demonstrate our approach has a much better performance compared to other baseline CNNs and state-of-the-art methods on various image datasets.

arXiv:2604.11913v1 Announce Type: new Abstract: Nutrition estimation of meals from visual data is an important problem for dietary monitoring and computational health, but existing approaches largely rely on single images of the finally completed dish. This setting is fundamentally limited because many nutritionally relevant ingredients and transformations, such as oils, sauces, and mixed components, become visually ambiguous after cooking, making accurate calorie and macronutrient estimation difficult. In this paper, we investigate whether the cooking process information from egocentric cooking videos can contribute to dish-level nutrition estimation. First, we further manually annotated the HD-EPIC dataset and established the first benchmark for video-based nutrition estimation. Most importantly, we propose V-Nutri, a staged framework that combines Nutrition5K-pretrained visual backbones with a lightweight fusion module that aggregates features from the final dish frame and cooking process keyframes extracted from the egocentric videos. V-Nutri also includes a cooking keyframes selection module, a VideoMamba-based event-detection model that targets ingredient-addition moments. Experiments on the HD-EPIC dataset show that process cues can provide complementary nutritional evidence, improving nutrition estimation under controlled conditions. Our results further indicate that the benefit of process keyframes depends strongly on backbone representation capacity and event detection quality. Our code and annotated dataset is available at https://github.com/K624-YCK/V-Nutri.

arXiv:2604.12056v1 Announce Type: new Abstract: Block-wise diffusion language models (DLMs) generate multiple tokens in any order, offering a promising alternative to the autoregressive decoding pipeline. However, they still remain bottlenecked by memory-bound attention in long-context scenarios. Naive sparse attention fails on DLMs due to a KV Inflation problem, where different queries select different prefix positions, making the union of accessed KV pages large. To address this, we observe that between consecutive denoising steps, only a small fraction of active tokens exhibit significant hidden-state changes, while the majority of stable tokens remain nearly constant. Based on this insight, we propose LOSA (Locality-aware Sparse Attention), which reuses cached prefix-attention results for stable tokens and applies sparse attention only to active tokens. This substantially shrinks the number of KV indices that must be loaded, yielding both higher speedup and higher accuracy. Across multiple block-wise DLMs and benchmarks, LOSA preserves near-dense accuracy while significantly improving efficiency, achieving up to +9 points in average accuracy at aggressive sparsity levels while maintaining 1.54x lower attention density. It also achieves up to 4.14x attention speedup on RTX A6000 GPUs, demonstrating the effectiveness of the proposed method.

arXiv:2604.11868v1 Announce Type: new Abstract: While medical Vision-Language models (VLMs) achieve strong performance on tasks such as tumor or organ segmentation and diagnosis prediction, their opaque latent representations limit clinical trust and the ability to explain predictions. Interpretability of these multimodal representations are therefore essential for the trustworthy clinical deployment of pretrained medical VLMs. However, current interpretability methods, such as gradient- or attention-based visualizations, are often limited to specific tasks such as classification. Moreover, they do not provide concept-level explanations derived from shared pretrained representations that can be reused across downstream tasks. We introduce MedConcept, a framework that uncovers latent medical concepts in a fully unsupervised manner and grounds them in clinically verifiable textual semantics. MedConcept identifies sparse neuron-level concept activations from pretrained VLM representations and translates them into pseudo-report-style summaries, enabling physician-level inspection of internal model reasoning. To address the lack of quantitative evaluation in concept-based interpretability, we introduce a quantitative semantic verification protocol that leverages an independent pretrained medical LLM as a frozen external evaluator to assess concept alignment with radiology reports. We define three concept scores, Aligned, Unaligned, and Uncertain, to quantify semantic support, contradiction, or ambiguity relative to radiology reports and use them exclusively for post hoc evaluation. These scores provide a quantitative baseline for assessing interpretability in medical VLMs. All codes, prompt and data to be released on acceptance. Ke

Discussion | Link

<p><strong>Release:</strong> <a href="https://github.com/datasette/datasette-ports/releases/tag/0.3">datasette-ports 0.3</a></p> <p>A small update for my tool for helping me figure out what all of the Datasette instances on my laptop are up to.</p> <blockquote> <ul> <li>Show working directory derived from each PID</li> <li>Show the full path to each database file</li> </ul> </blockquote> <p>Output now looks like this:</p> <pre><code>http://127.0.0.1:8007/ - v1.0a26 Directory: /Users/simon/dev/blog Databases: simonwillisonblog: /Users/simon/dev/blog/simonwillisonblog.db Plugins: datasette-llm datasette-secrets http://127.0.0.1:8001/ - v1.0a26 Directory: /Users/simon/dev/creatures Databases: creatures: /tmp/creatures.db </code></pre> <p>Tags: <a href="https://simonwillison.net/tags/datasette">datasette</a></p>

Midi Health CEO Joanna Strober set out to launch a menopause startup, but ended up building an AI company. Tammy Horton/Nikki Ritcher Women's health startup Midi Health recently passed a $1 billion valuation. To scale faster, it built a custom chatbot that avoids outdated menopause data. Midi Health CEO Joanna Strober shared how AI has changed how she operates her business. Midi Health CEO Joanna Strober originally set out to launch a menopause startup. Then, AI changed everything. Founded in 2021, Midi Health provides virtual care to menopausal women and recently crossed the $1 billion valuation mark in a February funding round. At Business Insider's The Long Play event in San Francisco on Tuesday, Strober said AI has transformed her business. "I thought I was starting a menopause company, but it turns out I'm building an AI company," Strober said onstage. Midi Health, based in Palo Alto, California, now uses AI to train thousands of its providers on how to answer patient questions. That's helped the company scale to serving more than 20,000 women a week — or millions a year. The move wasn't easy. Because research on women's health — especially menopause — is sparse, general-purpose AI chatbots trained on the open internet weren't good enough. Midi Health had to build its own chatbot using only high-quality data and excluding studies that have since been disproven. "There's a lot of old, outdated data," she said. AI dinners and office hours Beyond the chatbot, Midi Health also uses AI in its internal systems. Strober has been flying around the country to have dinner with the company's nurses, talking through how they should use AI at work. She said she introduced one nurse — who was trying to standardize hundreds of contracts — to Google Gemini. The job, which the nurse thought would take about a month, ended up taking 10 minutes. Midi Health has also started holding AI office hours, where the company's software engineers help different teams "AI-ify" themselves. Strober says the shift isn't leading to mass layoffs. "It's not threatening — it's really about augmenting their jobs," she said. Read the original article on Business Insider

One investor who has backed both companies told the FT that justifying OpenAI's recent round required assuming an IPO valuation of $1.2 trillion or more — making Anthropic's current $380 billion valuation look like the relative bargain.

Jason Blum shares how Blumhouse's original model has had to evolve with Hollywood. Tammy Horton/Nikki Ritcher Jason Blum said he got roasted when Blumhouse teamed up with Meta to make AI shorts. He said it taught him that AI won't be replacing filmmakers for a long time. Blum said he thinks online content creators have more to worry about than directors and writers. AI is a hot-button topic in Hollywood, but the modern king of horror was still surprised by the backlash he received when he teamed up with Meta. "We made three little AI shorts for Meta. We got destroyed on Twitter for doing that," Jason Blum, founder and CEO of Blumhouse Productions, told Peter Kafka at Business Insider's The Long Play event in San Francisco on Tuesday. "But I learned a lot." In October 2024, Meta announced it was partnering with Blumhouse to test its suite of AI models that can help with moviemaking, but Blum said the experience actually made him less concerned about AI replacing filmmakers. After doing those shorts, he said he was "very confident" that AI would not make better content for "a long, long time." There's been rapid improvements in generative AI, with uncanny AI clips of stars like Tom Cruise and Brad Pitt fighting, prompting sweeping declarations about the death of Hollywood. Blum said he thinks AI-generated content is more likely to compete with scrolling on social media, rather than watching a movie. "I think for once, creators have more to worry about than directors and writers," he said. Founded by Blum in 2000, Blumhouse has produced some of the most successful horror franchises of the 21st century, like "Paranormal Activity," "Insidious," and "The Purge" films, as well as Oscar-winning critical darlings like "Whiplash" and "Get Out." Its latest release, "Lee Cronin's The Mummy," hits theaters this Friday. Blum has been open about the inevitable impact of AI on Hollywood. He said Tuesday that while he would prefer AI in moviemaking to go away, he knows it's here to stay, comparing AI opposition to resisting technology. "It really bothers me when a lot of people in Hollywood just are like, 'We're not doing AI,'" he said, adding, "It ain't going away." Blum said he personally tries to experiment with AI and make it available as an option for filmmakers who are interested in using it, calling it a "personal choice." "If a director wants to use AI or a filmmaker or a producer or a writer wants to use AI, I want to make sure that my company can provide information about how to do that," he said. Read the original article on Business Insider

<p><strong><a href="https://ziglang.org/download/0.16.0/release-notes.html#Juicy-Main">Zig 0.16.0 release notes: &quot;Juicy Main&quot;</a></strong></p> Zig has <em>really good</em> release notes - comprehensive, detailed, and with relevant usage examples for each of the new features.</p> <p>Of particular note in the newly released Zig 0.16.0 is what they are calling "Juicy Main" - a dependency injection feature for your program's <code>main()</code> function where accepting a <code>process.Init</code> parameter grants access to a struct of useful properties:</p> <div class="highlight highlight-source-zig"><pre><span class="pl-k">const</span> <span class="pl-v">std</span> <span class="pl-k">=</span> <span class="pl-k">@import</span>(<span class="pl-s">"std"</span>); <span class="pl-k">pub</span> <span class="pl-k">fn</span> <span class="pl-en">main</span>(<span class="pl-v">init</span>: <span class="pl-k">std.process.Init</span>) <span class="pl-k">!</span><span class="pl-k">void</span> { <span class="pl-c">/// general purpose allocator for temporary heap allocations:</span> <span class="pl-k">const</span> <span class="pl-v">gpa</span> <span class="pl-k">=</span> <span class="pl-v">init</span>.<span class="pl-v">gpa</span>; <span class="pl-c">/// default Io implementation:</span> <span class="pl-k">const</span> <span class="pl-v">io</span> <span class="pl-k">=</span> <span class="pl-v">init</span>.<span class="pl-v">io</span>; <span class="pl-c">/// access to environment variables:</span> <span class="pl-v">std</span>.<span class="pl-v">log</span>.<span class="pl-v">info</span>(<span class="pl-s">"{d} env vars"</span>, .{<span class="pl-v">init</span>.<span class="pl-v">environ_map</span>.<span class="pl-v">count</span>()}); <span class="pl-c">/// access to CLI arguments</span> <span class="pl-k">const</span> <span class="pl-v">args</span> <span class="pl-k">=</span> <span class="pl-k">try</span> <span class="pl-v">init</span>.<span class="pl-v">minimal</span>.<span class="pl-v">args</span>.<span class="pl-v">toSlice</span>( <span class="pl-v">init</span>.<span class="pl-v">arena</span>.<span class="pl-v">allocator</span>() ); }</pre></div> <p><small></small>Via <a href="https://lobste.rs/s/4vvozb/zig_0_16_0_release_notes">Lobste.rs</a></small></p> <p>Tags: <a href="https://simonwillison.net/tags/zig">zig</a></p>

Discussion | Link

Jonathan Raa/NurPhoto via Getty Images In September, the Australian Recording Industry Association (ARIA) excluded catalogue music (recordings more than two years old) from the Australian bestseller single and album charts. From a marketing perspective this decision is logical, as it creates room to expose new recordings to the market. However, it also obscures the reality of the new music economy in Australia. My latest research – which looks at new music releases in Australia from 2000 to 2024 – shows a significant decline in the sale of new music since the adoption of music streaming. These findings point to a crisis for new and emerging artists in the Australian market. The new music market is shrinking In 2017, music streaming platforms, led by Spotify, became the dominant form of recorded music distribution in Australia. The shift from a purchase-based (CDs, vinyl and downloads) to an access-based (streaming) economy represented a fundamental change in the music business. Streaming platforms, with close to unlimited repertoire, enable and encourage passive listening via playlists and algorithmic recommendation. The result is that catalogue music has become the mainstay of the recorded music industry. From 2000 to 2018, new release music made up 99% of the ARIA annual top 100 singles, and 78% of the top 100 albums. But from 2022 to 2024, these figures dropped to 62% and 28%, respectively. The data indicate that since 2000, new music revenue in Australia has declined by 55% in actual and 71% in inflation-adjusted value. The rise of streaming has led to us spending more on music overall, but less of this is going to new music. My estimates suggest new music revenues in Australia have grown by just 4% since 2014, in a market that has doubled in value. A similar trend is evident overseas. In the United States, new music accounted for an estimated 65% of recorded music revenue in the pre-streaming economy, compared to 25–30% post-streaming. New talent can’t rely on industry From an Australian perspective, the challenges for new music have created concern about pathways for emerging artists, and the music industry’s commitment to developing them. Industry insiders I interviewed for the research highlighted how labels were playing a diminishing role in artist development. The stress on the new music economy – combined with the reduced presence of Australian artists in the ARIA charts – has led to ongoing calls for the government to support the industry via cultural policy initiatives. The recorded music industry also has a role to play in addressing the environment which it helped to create – particularly in regards to how artists are remunerated. The current “pro-rata” model used by streaming platforms places equal value on all streams, regardless of whether it is a catalogue track or new release. Under this model, there is no business incentive to prioritise new music. Adjusting this model, so that new releases are valued higher than catalogue music, could create this incentive. Major labels will likely resist change, as they reap the rewards of selling back catalogue at pure profit. But the idea of valuing new music over catalogue is not new. Before the streaming era, new release CDs were sold at full price and catalogue CDs were often sold at mid-price. This model reflected the costs associated with developing new products and provided business incentive by attaching greater profit margins to new releases. Prioritising long-term industry health Australia’s new music economy has experienced significant revenue decline and reduced industry commitment to new and emerging artists. For independent labels, which are largely dependent on new release revenue, the ability to compete in a catalogue market is limited. For artists, the lack of pathways to earning revenue may lead to disconnection with the sector. The role of Australian major record labels as generators of local talent is also in question. In addition to policy, a business incentive for record labels to invest in new music could enable the long-term health of the sector. Tim Kelly was an ARIA Chart Committee member from 2005 to 2017, and worked for the record labels Universal, Sony and Inertia from 1999 to 2017.

This week, Donald Trump posted an AI image of himself dressed in white robes, placing a glowing hand over an ill or deceased man in a hospital bed, as if to heal or resurrect him. The image, posted on Truth Social, was widely taken as him presenting himself as a Messianic Jesus figure. Sometime the next morning, he deleted the post. “I thought it was me as a doctor,” he explained to reporters, according to Time magazine. Jesus? “Only the fake news could come up with that one.” But the post was widely interpreted as blasphemous – including by conservative Catholic group CatholicVote.org. “I don’t know if the President thought he was being funny or if he is under the influence of some substance or what possible explanation he could have for this OUTRAGEOUS blasphemy,” declared Megan Basham, a prominent conservative Protestant Christian writer, on X. “But he needs to take this down immediately and ask for forgiveness from the American people and then from God.” “I was very grateful to see how many conservative Christians immediately denounced the blasphemous Jesus/Trump image,” said pastor Doug Wilson, who recently led a prayer service at the Pentagon and founded the network of churches War Secretary Pete Hegseth belongs to. What is blasphemy? Within the Christian tradition, blasphemy has historically been an unstable, shifting idea. But, simply put, it means speech, thought or action that shows contempt for – or mockery of – God and sacred matters. Judaism and Christianity’s concept of blasphemy came from the injunction in the Old Testament not to revile God. Within the Old Testament, it was treated as a crime, punishable by death: “One who blasphemes the name of the Lord shall be put to death; the whole congregation shall stone the blasphemer.” The New Testament expanded the concept to include the rejection of Jesus. Eventually, cursing, reproaching, challenging, mocking, rejecting or denying Jesus became blasphemous. More particularly, posing as Jesus or asserting powers that belong only to him was considered blasphemous in medieval times. The “Christs” that emerged were treated harshly, as dangerous heretics. This is where Trump’s presentation of himself as Jesus would undoubtedly be considered blasphemous. Broadly, anything said or done that offended believers could be construed as blasphemous. Catholics at the time of the Reformation in the 16th century tended to brand those who offended them as heretics. Protestants generally preferred the term blasphemy for anything they disliked or disagreed with. For example, 16th century theologian Martin Luther – a seminal figure in the Protestant Reformation – condemned as blasphemous not only Catholics, but also Jews and Muslims. Blasphemy as sin or crime James Nayler. Britannica From the 17th century onwards, blasphemy became not so much an offence against God as one against society. Within the unstable societies of early modern Europe, blasphemy was viewed as socially and politically subversive and prosecuted as such. The Quaker James Naylor was imprisoned in 1656 for reenacting Jesus’s entry into Jerusalem on Palm Sunday. Early in the 17th century, blasphemy crossed to the United States. Virginia’s first law code specified death for blaspheming the divine Trinity. Despite the first amendment to the US constitution, protecting free speech, blasphemy laws were regularly enacted. The US Supreme Court didn’t rule that laws against blasphemy infringed the right to free speech until after World War II. Several states still have blasphemy on their books. England’s Blasphemy Act of 1697, which criminalised the denial of the Holy Trinity, the truth of Christianity, or the divine authority of the Bible, carried over into the colonies of Australia and New Zealand. Blasphemy is no longer an offence under Australian federal law, though laws governing it vary across the states: it’s still in the criminal code of many of them. New Zealand’s criminal code deals with “blasphemous libel” as part of “crimes against religion, morality and public welfare”. Is there blasphemy in Islam? Within Islam, there is no exact equivalent to “blasphemy”. But the idea of the “word of infidelity” is analogous to it. In practice, it amounts to mockery of God, the prophet, or the Islamic tradition generally. So, when Trump mockingly declared “Praise be to Allah” in a recent post, he was guilty of blasphemy in Muslim eyes. Conservative commentator Dinesh D'Souza compared it to the Old Testament account of the prophet Elijah, who mocked the prophets of Baal in 1 Kings 18. The Islamic advocacy group Council on American-Islamic Relations called it “disturbing” and “offensive to Muslims”. Laws against blasphemy are actively enforced in many modern Islamic states. Does blasphemy matter? It is not blasphemous to speak or publish opinions that are hostile to Christianity, Judaism or Islam – or for that matter, any religion. What matters is not so much the substance of criticism as the manner in which it is made. We should only worry when criticism becomes a form of “religious hate speech”. The question we should ask is about intent. In a secular society, where we do identify ill intent, we may wish to think about “blasphemy” as a matter of public morality, not theology. So, what about Donald Trump’s post? Does it matter? If we consider “blasphemy” to include the mocking of religion, there is little doubt that Trump’s mockery of Islam is blasphemous. If we believe his deleted Truth Social post was intended to suggest he is Jesus – or in some sense divine – then Christians are entitled to consider him blasphemous. That said, from a secular perspective, it is more self-indulgent foolishness than hate speech – but nonetheless, extremely inappropriate for a US president. Philip C. Almond 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.

Declan Sun/Unsplash The explosion of generative artificial intelligence (AI) tools has provoked both hopes and anxieties about the potential benefits and harms of this technology. In advanced economies, people are almost equally worried and optimistic about it. This is perhaps unsurprising. AI consumes vast amounts of natural resources yet promises to save the planet. It may improve human efficiency and productivity, while putting millions out of work. For many white-collar workers, AI use now seems non-optional. The messaging is clear – get on board or be left behind. Amid this uncertainty and rapid technological uptake, concerned citizens have made efforts to resist AI. One form of AI resistance, aimed at sabotaging the functionality of AI large language models, is data poisoning. But how accessible is it to the everyday person? And what is at stake in its use? What is AI resistance? Acts of AI resistance range from social sanctions and boycotts, to strikes, protests, public outcry and lawsuits. Driving these acts are perceived threats to jobs, ethics, safety, democracy and sovereignty, and the environment. AI is also described as an existential risk to creative industries, including music, fiction and film. In the United Kingdom, generative AI has been characterised as an “industrial scale theft” that threatens a £124.6 billion (A$237bn) creative sector and more than 2.4 million jobs. People have long used civil disobedience to address social injustices. Famously, Rosa Parks’ refusal to sit at the back of a bus in Alabama led to a 13-month bus boycott by tens of thousands of Black residents. It only ended when racial segregation on public transport was deemed unconstitutional in the United States. Acts of sabotage have also long been central to collective action against injustice. In fights for labour rights, workers have employed diverse tactics to reduce efficiency and productivity. This has ranged from hotel workers putting salt in sugar bowls to farm workers breaking machinery. Data poisoning can be viewed as a modern version of these historic actions. How does data poisoning work? Data poisoning means deliberately inserting misleading, biased, or nonsensical content into the data AI models learn from, to make their outputs worse. Only 250 poisoned documents in a dataset could compromise outputs across AI models of any size. There are various ways to poison data. Some require highly technical skills, others are accessible to anyone with an internet connection – if their text or images are used as training data. Researchers have developed several data poisoning tools that exploit the vulnerabilities of AI models. Glaze and Nightshade enable artists to make poisoned visual images that can’t be used as training data. The tool CoProtector defends against the exploitation of open source code repositories like Github. Monash University and the Australian Federal Police have created Silverer, enabling social media users to doctor personal images to prevent them from being used in deepfakes. Example images of AI model output generated with data poisoned with the Nightshade tool. Shan et al., arXiv (2023), CC BY But you don’t need a tool or advanced skills to affect AI. Simply creating websites with factitious information, making jokes in Reddit, feeding models their own outputs, or editing Wikipedia can poison data. Data poisoning is commonly presented as a dangerous act perpetrated by “cyber criminals” or “malicious actors”. But what if it’s used to protect human rights? Is data poisoning legal? Is it ethical? Legal obligations related to data poisoning are often directed to AI developers and organisations. The EU Artificial Intelligence Act requires that appropriate measures are adopted to prevent and detect data poisoning. The legal status of AI data poisoning by individual users is less clear. Criminal penalties may apply under US or UK computer fraud and misuse laws. Interference with an AI model would also likely breach the terms of service of AI companies. If AI data poisoning is unlawful, questions could still be asked about its ethical status. Philosophers have long recognised that civil disobedience can be justifiable in circumstances where legally sanctioned practices produce serious injustice. If AI companies are operating with state approval in ways that impact citizens’ rights to privacy, copyright, safe and secure work, quality education, social and sexual safety, data poisoning may constitute ethical civil disobedience. For philosopher John Rawls, “[civil disobedience] is one of the stabilising devices of a constitutional system, although by definition an illegal one”. If the intention is to prevent mass unemployment, preserve the integrity of elections, and protect against social harms (suicide, child abuse, increased human isolation, loss of human creativity and environmental degradation), data poisoning could align with the principles of justice that underpin democratic social institutions. A significant problem with data poisoning is that even if models become compromised – and outputs grow inconsistent, misleading, or nonsensical – users overly trust AI systems. Data poisoning then could contribute to harms it seeks to resist, amplifying the inaccuracy of systems humans are increasingly relying on, irrespective of their quality and effects. Data poisoning is not simply an immoral cyber crime. It can be an ethically complex strategy to address social injustices. AI development needs to be of collective benefit and aligned with public values and interests. If AI company employees are asking “Are we the baddies?”, history may prove that in some cases data poisoners are on the side of good. The authors do not work for, consult, own shares in or receive funding from any company or organisation that would benefit from this article, and have disclosed no relevant affiliations beyond their academic appointment.

The deal aims to accelerate AI adoption, train workers, and develop cybersecurity partnerships — the latest move by a hyperscaler to compete for sovereign AI and data centers.

Using importance sampling with fine-tuned donor prefills to predict reward hacking emergence during training

If you have never tried a video model before, now is the time.

<p><strong><a href="https://github.com/simonw/datasette/pull/2689">datasette PR #2689: Replace token-based CSRF with Sec-Fetch-Site header protection</a></strong></p> Datasette has long protected against CSRF attacks using CSRF tokens, implemented using my <a href="https://github.com/simonw/asgi-csrf">asgi-csrf</a> Python library. These are something of a pain to work with - you need to scatter forms in templates with <code>&lt;input type="hidden" name="csrftoken" value="{{ csrftoken() }}"&gt;</code> lines and then selectively disable CSRF protection for APIs that are intended to be called from outside the browser.</p> <p>I've been following Filippo Valsorda's research here with interest, described in <a href="https://words.filippo.io/csrf/">this detailed essay from August 2025</a> and shipped <a href="https://tip.golang.org/doc/go1.25#nethttppkgnethttp">as part of Go 1.25</a> that same month.</p> <p>I've now landed the same change in Datasette. Here's the PR description - Claude Code did much of the work (across 10 commits, closely guided by me and cross-reviewed by GPT-5.4) but I've decided to start writing these PR descriptions by hand, partly to make them more concise and also as an exercise in keeping myself honest.</p> <blockquote> <ul> <li>New CSRF protection middleware inspired by Go 1.25 and <a href="https://words.filippo.io/csrf/">this research</a> by Filippo Valsorda. This replaces the old CSRF token based protection.</li> <li>Removes all instances of <code>&lt;input type="hidden" name="csrftoken" value="{{ csrftoken() }}"&gt;</code> in the templates - they are no longer needed.</li> <li>Removes the <code>def skip_csrf(datasette, scope):</code> plugin hook defined in <code>datasette/hookspecs.py</code> and its documentation and tests.</li> <li>Updated <a href="https://docs.datasette.io/en/latest/internals.html#csrf-protection">CSRF protection documentation</a> to describe the new approach.</li> <li>Upgrade guide now <a href="https://docs.datasette.io/en/latest/upgrade_guide.html#csrf-protection-is-now-header-based">describes the CSRF change</a>.</li> </ul> </blockquote> <p>Tags: <a href="https://simonwillison.net/tags/csrf">csrf</a>, <a href="https://simonwillison.net/tags/security">security</a>, <a href="https://simonwillison.net/tags/datasette">datasette</a>, <a href="https://simonwillison.net/tags/ai-assisted-programming">ai-assisted-programming</a></p>

Dario Amodei is the CEO and cofounder of Anthropic. Michael M. Santiago/Getty Images Anthropic has recently received multiple offers from VCs to invest at valuations as high as $800 billion. Buzzy startups frequently field preemptive offers from investors only to rebuff them. Still, the valuations indicate the feverish demand for Anthropic shares that has overtaken Silicon Valley. Anthropic has fielded multiple offers from VCs valuing the startup behind Claude at as much as $800 billion in recent weeks, more than double its current valuation, according to multiple people familiar with the matter. Buzzy startups frequently field preemptive offers from investors only to rebuff them. Still, the valuations VCs are floating indicate the feverish demand that has overtaken Silicon Valley for a stake in Anthropic as the company prepares for a possible IPO later this year. A spokesperson for Anthropic declined to comment. Anthropic closed a funding round in February led by GIC and Coatue that valued the company at $380 billion. That is less than half the $852 billion valuation OpenAI achieved in the funding round it closed last month. Anthropic is valued at $688 billion on Caplight, a secondary exchange where investors can trade shares of privately held companies. That is up 75% in three months. Investors and founders have been wowed by Anthropic's torrid growth and momentum around AI-powered coding assistant Claude code. "They're crushing it," Jared Quincy Davis, founder and CEO of Mithril, an AI cloud platform, said about Anthropic last week at HumanX, an AI conference. Anthropic's run-rate revenue, the revenue a company expects to collect over a year, has soared to $30 billion, up from $9 billion at the end of last year, the company announced last week. It also added that over 1,000 business customers are spending more than $1 million a year, a figure that has doubled in less than two months. Last week, Anthropic announced its latest model, Mythos, which it said is so powerful it cannot yet be unleashed upon the general public because of the risk of cyber attacks. "The Mythos model is a huge deal, Tomasz Tunguz, founder and general partner of Theory Ventures, said last week at HumanX. "There's a tremendous amount of excitement." Read the original article on Business Insider