Paleontologists studying a fossil unearthed in southern Manitoba, Canada, have identified what they believe are two bite marks on the internal skeleton of the ancient giant cephalopod Enchoteuthis.
The post Fossil of Giant Squid-Like Creature from Cretaceous Period Bears Marks of Deadly Attack appeared first on Sci.News: Breaking Science News.
Anthropic will embed watermarks in the text and files generated by future models it launches in the EU, as part of its effort to comply with content and transparency rules in the bloc’s AI Act.
[…]
The move may further amplify the appeal of open weight models and alienate Claude customers, who don’t necessarily want consumers of their AI-generated content to know its provenance.
[…]
Anthropic itself is already hedging about the utility of its marking method, noting that detected marks are not conclusive evidence that Claude produced the content and that the absence of marks cannot guarantee that AI wasn’t involved in the creation of a particular piece of content.
Claude models launched in the EU on or after August 2, 2026 will support machine-readable marking at launch. Generated text will carry embedded watermarks, and generated files will include digitally signed provenance metadata where supported.
[…]
Marks will apply to output from supported Claude models across Claude Platform (API), Claude, Claude Code, Claude Cowork, and Claude Tag, and wherever Claude is offered, worldwide. Some platforms or features may not support certain marking types.
[…]
We’ll support users and other third parties to detect Claude’s marks, as the Code requires, and we’ll share details in forthcoming documentation.
Anthropic is claiming this is for compliance with an EU law but that “Marking will apply to output from supported models wherever Claude is offered, worldwide.”
This is infuriatingly opaque. I won’t speak to the “signed provenance metadata” they say they’ll be embedding in generated files like PDFs, SVGs, PNGs, and JPEGs. That’s important but it’s complex. Plain text is simple. A string of text is just one character followed by another. I presume that Claude is going to start “embedding” invisible characters/byte sequences between visible characters? What they’re claiming to do here seems impossible, frankly, and anything they attempt to embed invisibly is going to cause immediate problems.
[…]
If the “watermark” is not comprised of invisible characters but rather visible ones, how in the world does this jibe with their claim that it’s “imperceptible”? (This Reddit thread claims that’s how it will work — “It’s a form of steganography where the model subtly biases its word choice to create a statistical pattern that can be detected later.” How in the world can that be squared with “it doesn’t change the meaning, quality, or readability”?) And any sort of semantic detection like this is going to cause false-positive problems.
Previously:
Update (2026-08-17): Anthropic provides more details:
Watermarking won’t be specific to Claude. As of August 2, the EU requires AI providers serving its market to mark AI-generated content. Other major model developers have signed the same Code of Practice and will be implementing their own watermarks.
[…]
In internal testing, we’ve seen no impact of watermarking on the content, level of creativity, or readability of Claude’s text. In the SynthID-Text paper, which introduced the technique we use, Google DeepMind tested this impact by serving a model that used watermarking to a portion of their Gemini traffic and comparing thumbs-up and thumbs-down ratings. They found no statistically significant differences from the unwatermarked model. And in a controlled study, human raters comparing watermarked and unwatermarked answers side-by-side saw no difference in quality.
[…]
There are limitations to the effectiveness of watermarking. Using our key, one can only answer the question “What is the likelihood this was partly written by Claude?” It doesn’t confirm whether the text was human-written, and it can’t tell whether the text was written by a different AI (even if that other AI uses watermarking, it would have a different key; it might also use a different watermarking method altogether). Detecting a watermark also doesn’t work well on small samples, where there are fewer word choices and thus less information to go on. As a passage increases in length, confidence about Claude’s involvement increases too.
Watermarking is sparser on factual passages where there are fewer choices that can be made without decreasing the accuracy of the text. […] For the same reason, code—which in very many cases has to be exact—has generally less watermarking than some other forms of text.
John Gruber (Mastodon, Bluesky, Hacker News):
I initially guessed “invisible characters” not because I didn’t think of the semantic word-choice technique, but because I was a fool who took Anthropic at its word in their description of what they would do.
[…]
They say “imperceptible” and “doesn’t change the meaning, quality, or readability”. Their words. Not almost imperceptible. Not slightly changes the meaning, quality, or readability.
There seems to be a slight of hand here, where Google studied whether humans saw a difference in quality but then Anthropic claimed that there was no difference in meaning. Changing words will almost always change the meaning. Perhaps not in an important way, but a change is a change. Pretending otherwise is offensive to a writer. The response to this line of argument seems to be that LLMs already have randomness and that the marking changes will be within the level noise that’s already there. To me, this presents a pessimistic view of the capabilities of LLMs, contrary to Anthropic’s normal claims. The Google study was published in 2024 and didn’t test Claude. The models have changed a lot since then, but they are assuming that the precision hasn’t advanced enough to be affected, nor will it in the future. Or that when there is a tension between watermarking and quality they will pick the former. And perhaps you won’t be able to tell because everyone else will be watermarking, too.
But the very best description of the general idea behind the technique is an interactive essay by James Padolsey, “How AI Text Watermarking Works”.
[…]
This isn’t just about text one might generate with the intention of passing it off as their own natural work. This isn’t even about LLM proofreading of work written by hand. Anthropic is saying that all new Claude models are going to adulterate every single bit of text longer than 200 tokens (~150 words) they generate, including everything it presents to its users to read. So even in a private conversation between a user and Claude, which will never be read by anyone other than the user, Claude will begin making word choices in the name of marking its output in statistically predictable ways rather than maximizing clarity and precision.
[…]
Taken literally, compliant LLM terms of service must forbid users from rephrasing the output from models that comply with this regulation, because the word choices are the marks. But it’s not the European Union that is trying to impose their absurd, impractical, witch-hunt-fueling regulation on the entire world. That falls on Anthropic.
Complying with this, particularly with regard to text, is only going to create problems for honest users. Dishonest users attempting to pass off AI-generated text as their own writing (students, employees, whoever) will simply circumvent detection through non-compliant AI paraphrasing tools.
I don’t love that checking for a watermark requires submitting the text to the AI provider, and all you get back is an opaque result with no way to verify it.
James Padolsey (via John Gruber):
The move is Anthropic’s response to Article 50(2) of the EU AI Act, which requires providers to ensure that outputs are “marked in a machine-readable format and detectable as artificially generated or manipulated”.
[…]
The Act exempts standard editing, but many legitimate assistive uses require more substantial rewriting while leaving the ideas, judgment and responsibility with the human. The same thought that led to this law could have applied to calculators at the time of their inception, had their outputs revealed themselves through artefacts. Thankfully, a sum borne of the brain is treated no differently from one produced by a calculator. Likewise with spellcheckers. To make assistance suspect only once the tool becomes capable enough to compose a whole sentence is not a principled boundary. It is a moral premium placed on difficulty itself.
Anthropic has nevertheless chosen a blanket, model-level implementation that appears broader than the law’s minimum requirement. That may be convenient compliance engineering, but it discards distinctions the law expressly attempted to preserve. The result is a signal broad enough to implicate harmless and assistive use, yet fragile enough to be removed by a motivated person through substantial recomposition. It risks concentrating suspicion on ordinary and assistive users while remaining weakest against deliberate deception.
Update (2026-08-20): See also:
Without getting into the weeds, I still think Anthropic’s claim of their being no change in the meaning, quality, or readability is not accurate, using the plain meanings of those terms. If you read the papers, there are ways of ensuring that the frequencies of individual tokens or sequences of them are not affected in aggregate. But this is a different claim, and Anthropic does not specify which tradeoffs they’ve chosen there. I’m willing to grant that they can make the effect very small.
At this point, I think the most interesting question is, what will be the effects of these watermarks?
AI watermarking is security theater. There is absolutely nothing to prevent people from removing any possible watermark in text, images, video or other data. There is no way to prevent false positives, or false negatives, in detectors, either. It’s all superficial nonsense.
I find it interesting to be able to test a particular passage of text, but what am I supposed to do with that information? Even without watermark removal, a negative for a given model doesn’t tell you whether another model was used. And I’m uncomfortable with convicting someone based on a watermark given that false positives are possible and independent verification is not.
I asked Gemini about the benefits, and it said:
Regulatory Compliance and Policy
- Legal adherence: Helps major tech platforms satisfy emerging mandates requiring automated systems to label machine-generated media.
- Ecosystem tracking: Gives platforms and enterprises a high-level telemetry signal to monitor the spread of synthetic content across public channels.
Triage and Risk Mitigation
- Mass moderation: Assists automated filters in flagging bulk, unvetted bot content or spam networks where statistical trends matter more than single-file perfection.
- Provenance auditing: Provides an extra layer of metadata for digital media archives, helping investigators spot obvious clusters of AI-generated assets.
In other words, it’s not so much about individual pieces of content.
Update (2026-08-24): John Gruber:
Also, more and more, I do think this has little to do with the EU regulation and everything to do with Anthropic and Google wanting to watermark the text that Claude and Gemini generate for their own purposes, to identify text to avoid training future models on to avoid model collapse/AI inbreeding.
See also: Accidental Tech Podcast.
Google’s upcoming Pixel 11 Pro XL has reportedly appeared on Geekbench 7, revealing early performance details of the company’s next-generation Tensor G6 chipset. While the new processor is expected to bring improvements in efficiency and AI capabilities, the leaked benchmark results suggest that raw performance gains may be relatively modest compared with the previous-generation Tensor G5.
According to Igeekphone, the Pixel 11 Pro XL and other Pixel 11 series models are expected to be unveiled at Google’s annual Made by Google event on August 12. Ahead of the official announcement, benchmark information has provided the first look at the performance of Google’s upcoming flagship processor.
The Geekbench 7 listing reportedly shows that the Google Pixel 11 Pro XL achieved:
These results represent the first known real-world benchmark scores for the Tensor G6 chipset, which has previously appeared only through leaks and industry reports.
The benchmark listing also confirms that the tested device is equipped with:
While benchmark results do not reveal every aspect of real-world performance, they provide an early indication of the processing capabilities of Google’s next flagship smartphone platform.
According to the Geekbench database information, the Tensor G6 appears to use a seven-core CPU architecture consisting of three core clusters.
The reported configuration includes:
The listing does not confirm whether the Tensor G6 is manufactured using a 2nm process technology, despite previous reports suggesting that Google could become one of the first smartphone manufacturers to adopt a 2nm Tensor chip.
Although the Tensor G6 represents a new generation of Google’s custom smartphone chipset, the leaked benchmark results suggest that the performance improvement may not be dramatic.
For comparison, the previous-generation Pixel 10 Pro’s Tensor G5 reportedly achieved:
Compared with those numbers, the Tensor G6 appears to offer a small increase in single-core performance but a lower multi-core score in this early Geekbench test.
If the leaked results are accurate, the Pixel 11 Pro XL may not deliver a major upgrade in traditional computing performance. Instead, Google may continue focusing on areas where Tensor chips have historically stood out, such as artificial intelligence processing, photography enhancements, and software optimization.
The Pixel 11 series has previously appeared in Geekbench databases as well.
Earlier testing of the standard Pixel 11 model reportedly showed:
That listing also indicated a seven-core Tensor G6 architecture, although the performance results were significantly lower than the Pixel 11 Pro XL benchmark.
Differences between models could be related to software optimization, testing conditions, clock speeds, or hardware configurations.
Google’s Tensor processors have traditionally prioritized artificial intelligence and machine learning capabilities rather than competing directly with flagship chips from Qualcomm and Apple in raw benchmark performance.
The Tensor G6 is expected to bring improvements in areas such as:
For Pixel users, these software-driven improvements may be more noticeable in daily use than benchmark score increases.
Google is expected to officially introduce the Pixel 11 series along with the Pixel Watch 5 during the upcoming Made by Google event on August 12.
The Pixel 11 lineup is expected to include multiple models, including:
The Pixel 11 Pro XL is expected to serve as the top-tier model, featuring the largest display, highest memory configuration, and the most advanced camera system in the lineup.
The leaked Geekbench results for the Google Pixel 11 Pro XL suggest that the upcoming Tensor G6 chipset may deliver only moderate improvements in traditional processing power. While the move to a newer architecture and potential 2nm manufacturing process could improve efficiency, the benchmark numbers indicate that Google is not necessarily chasing the highest raw performance scores in the smartphone market.
Instead, the Pixel 11 series is likely to continue Google’s strategy of combining custom hardware with advanced AI software experiences. For potential buyers, the most important upgrades may come from new AI features, camera improvements, and software innovations rather than a major leap in benchmark performance.
The post Google Pixel 11 Pro XL Geekbench Scores Leak: Tensor G6 Shows Limited Performance Gains appeared first on IGeeKphone China Phone, Tablet PC, VR, RC Drone News, Reviews.
Um Apple I funcional, de 1977, entrou em leilão pela RR Auction. E o destaque vai para o seu valor estimado, de US$300 mil — os lances começam em US$10 mil.
Ele é um dos 50 modelos do segundo lote produzido pela Apple e está instalado em uma mala adaptada a partir de uma máquina de escrever Smith-Corona, sendo restaurado pelo especialista Corey Cohen, em junho, para funcionamento.
O modelo do leilão, “Neumark”, destaca-se por incluir mais do que a placa de circuito montada, a única coisa que os compradores do Apple I recebiam ao adquirir a máquina.
O item faz parte da segunda parte do “Steve Jobs & the Computer Revolution: The Apple 50th Anniversary Auction”, um leilão comemorativo que inclui objetos históricos da Apple, como um iPhone original lacrado, um protótipo do primeiro iPod — cada um podendo chegar a um valor de US$50 mil —, documentos como um cartão de visitas da Maçã assinado por Steve Jobs, um cheque da Apple assinado por ele (destinado a Elizabeth Holmes, primeira responsável pela contabilidade da empresa) e, ainda, itens pessoais do cofundador e ex-CEO 1Chief executive officer, ou diretor executivo. da Maçã.
Os lances poderão ser feitos até o dia 20 de agosto, às 18h (horário da Costa Leste dos Estados Unidos). Às 19h, entra em vigor a regra de prorrogação automática do leilão, que estende-o em intervalos de 30 minutos a partir do registro de novos lances.
via AppleInsider
From Nevada to Manchester, developers are trialling innovative solutions to clean energy’s biggest challenge
In the deserts of the United Arab Emirates a sprawling clean energy project, stretching across an area roughly the size of 12,600 football fields, will play host to a breakthrough allowing solar energy to power the equivalent of half a million homes through the night.
The Gulf state has been steadily combining 5.2GW of solar power capacity with 19GWh of battery storage to create the largest battery scheme in the world.
Continue reading...
© Photograph: UCG/Universal Images Group/Getty Images
© Photograph: UCG/Universal Images Group/Getty Images
© Photograph: UCG/Universal Images Group/Getty Images
Thousands of fossilized bones from a Cretaceous-period bonebed in Wyoming, the United States, offer rare physical evidence that Tyrannosaurus rex fed on the carcasses of a duck-billed dinosaur species called Edmontosaurus annectens.
The post Tyrannosaurus rex Scavenged Duck-Billed Dinosaurs in Ancient Wyoming, Bite Marks Reveal appeared first on Sci.News: Breaking Science News.
In a groundbreaking advancement for artificial intelligence in healthcare, researchers at Mass General Brigham have unveiled BRIDGE, a comprehensive multilingual benchmark designed to critically assess how effectively large language models (LLMs) comprehend and interpret clinical patient-care text. Unlike existing AI evaluation metrics that rely primarily on structured and standardized medical exam questions, BRIDGE is engineered to engage with the multifaceted and complex language found in real-world clinical communications, including electronic health records (EHRs), case reports, and patient-doctor consultations, across nine different languages. This innovation represents a transformative step toward making AI tools more reliable and contextually aware in actual healthcare settings.
The existing paradigm in medical AI evaluation has predominantly centered on licensing exam questions composed in a controlled, formalized medical lexicon. These assessments, while rigorous, often fall short of reflecting the nuanced, variable, and sometimes ambiguous language used in actual clinical environments. The BRIDGE benchmark circumvents this limitation by employing authentic clinical texts, which capture the complexities and heterogeneity inherent to patient care dialogues, medical documentation, and clinical decision-making processes. This shift brings essential granularity and relevance to model performance metrics, providing clinicians and developers with more actionable insights.
The creators of BRIDGE demonstrated the stark contrast in performance between conventional exam-based evaluations and real-world clinical comprehension. For instance, the highest-performing LLM evaluated scored impressively on standard medical exams, achieving up to 92%. However, when subjected to BRIDGE’s rigorous clinical text test, the same model’s proficiency plummeted to only 44.8%. This significant disparity exposes considerable gaps in the AI’s ability to grasp the subtle clinical context, implicit meanings, and domain-specific language patterns prevalent in healthcare communications.
To validate the robustness and breadth of BRIDGE, the research team conducted a systematic performance evaluation of 95 distinct LLMs sourced from 59 different clinical AI initiatives. These models were subjected to a comprehensive battery of real-world clinical tasks encompassing the entire patient care continuum, ranging across 14 medical specialties. The tasks included patient triage, extraction of critical information from records, diagnostic reasoning, prognostic forecasting, and administrative functions such as billing code assignment. This extensive benchmarking provides a panoramic view of LLM capabilities and limitations in diverse clinical scenarios.
One of the more innovative features of BRIDGE is its public and continuously updated leaderboard hosted on the Hugging Face platform. This dynamic leaderboard catalogs the performance metrics of over 100 LLMs, enabling stakeholders including clinicians, health administrators, and AI developers to track comparative model efficacy in near real-time. The leaderboard thus fosters transparency and spurs iterative improvements by highlighting strengths and vulnerabilities within specific clinical tasks or language domains.
Another salient discovery made possible by BRIDGE is the identification of variability in AI performance across different medical specialties. Given that the benchmark corpus includes nine languages, the tool also illuminates disparities in model effectiveness when dealing with non-English clinical texts. This multilingual adaptability is particularly crucial as healthcare becomes more globally interconnected, underscoring the urgent need to develop culturally and linguistically sensitive AI applications that avoid exacerbating health inequities.
The scientific rigor of BRIDGE is underscored by its deep collaboration among experts spanning pharmacoepidemiology, pharmacoeconomics, clinical medicine, and computational modeling. The team includes senior authors such as Jie Yang, PhD, and Joshua Lin, MD, along with co-first authors Jiageng Wu and Bowen Gu, whose collective expertise was critical in ensuring the benchmark’s relevance and accuracy. Such interdisciplinary engagement is vital for bridging the gap between AI innovation and clinical applicability.
BRIDGE’s architecture and methodology leverage advanced computational simulation and modeling techniques, facilitating nuanced task designs that mimic real clinical workflows. This approach allows the benchmark to capture the dynamic and context-rich nature of clinical text interactions, incorporating elements like colloquial doctor-patient exchanges, complex diagnostic narratives, and procedural documentation. Consequently, BRIDGE functions as a high-fidelity proxy for real healthcare communication scenarios, offering a much-needed calibration tool for medical LLMs.
Funding generously provided by the Patient-Centered Outcomes Research Institute, the National Institutes of Health, and institutional scholarships reflects the strategic priority placed on refining AI’s role in healthcare delivery. Moreover, the rigorous conflict of interest disclosures and adherence to institutional compliance underscore the study’s commitment to transparency and ethical research standards. These factors enhance the credibility of BRIDGE as a benchmark tool destined to influence clinical AI development profoundly.
Importantly, BRIDGE is more than a passive evaluation tool—it is a catalyst for elevated AI design tailored to the clinical domain. By exposing the blind spots and differential performance across specialties and languages, it empowers AI developers to iterate more purposefully, embedding clinical nuance and real-world complexity into model training. This iterative feedback loop has the potential to accelerate the maturation of AI models from theoretical capabilities to practical clinical decision-support systems.
The release of BRIDGE is poised to address one of the persistent challenges in clinical AI—trust. Reliable understanding of patient-care language is paramount to fostering clinician confidence in AI-assisted diagnoses, prognoses, and patient management recommendations. The benchmark’s capacity to expose and rectify shortcomings before clinical deployment mitigates risks of errors caused by misinterpretation of nuanced clinical text, thereby safeguarding patient safety and improving care outcomes.
In closing, BRIDGE exemplifies a paradigm shift that acknowledges the inherent complexities of clinical language and seeks to elevate the fidelity of AI’s interpretative functions accordingly. As healthcare continues its digital transformation, integrating intelligent systems into everyday practice demands benchmarks as sophisticated and reflective as the environments they serve. BRIDGE sets a new gold standard in this endeavor, bridging the divide between cutting-edge AI performance and meaningful clinical utility.
Subject of Research: People
Article Title: BRIDGE: benchmarking large language models for understanding real-world clinical practice texts
News Publication Date: 17-Jun-2026
Web References:
References: Wu, J. et al. “BRIDGE: Benchmarking Large Language Models for Understanding Real-world Clinical Practice Text” Nature Biomedical Engineering DOI: 10.1038/s41551-026-01719-2
Keywords: Artificial intelligence, machine learning, clinical medicine, large language models, electronic health records, multilingual AI, clinical text comprehension, medical AI benchmarking, healthcare AI, patient care AI
© Loren Elliott for The New York Times
Login