AI News Roundup – June 28, 2025

June 28, 2025 – The world of artificial intelligence saw a flurry of developments this week, spanning major corporate moves, cutting-edge research breakthroughs, new AI-powered tools, as well as regulatory and ethical milestones. From tech giants recruiting top AI talent and investing in infrastructure, to breakthroughs in healthcare and robotics, and new laws beginning to shape AI’s future – here are the key updates in AI that everyone is talking about.

Corporate AI Moves and Investments

Meta Recruits OpenAI Talent and Boosts AI Investments: Facebook parent Meta made headlines by hiring Trapit Bansal, a key researcher behind OpenAI’s advanced reasoning model cointelegraph.com. Bansal joins several other former OpenAI scientists recently recruited by Meta cointelegraph.com, part of CEO Mark Zuckerberg’s push to strengthen Meta’s AI capabilities. The company aims to train its AI systems on more real-world data to improve their reasoning and planning skills cointelegraph.com. Alongside the talent grab, Meta has been pouring money into AI infrastructure – in June it acquired a 49% stake in data-labeling firm Scale AI (valuing Scale at nearly $15 billion) cointelegraph.com. Meta also secured a 20-year supply of nuclear power (1.1 GW from Constellation Energy) to fuel its AI data centers starting in 2027 cointelegraph.com, underscoring the immense energy needs of AI at scale. In the defense arena, Meta partnered with military tech firm Anduril to develop AI-powered augmented reality headsets for the U.S. military cointelegraph.com – a project integrating Anduril’s battlefield data platform into Meta’s AR devices for soldiers. All these moves signal Meta’s determination to be at the forefront of the “AI race,” backed by significant talent and resources.

Amazon’s Quiet AI Expansion: While flashier AI firms grabbed headlines, Amazon’s steady investment in AI has turned it into an understated winner. The company’s stock price has nearly doubled in the past three years, a rise analysts attribute in part to AI-driven growth across its businesses binaryverseai.com. In an analysis for Nasdaq, Jennifer Saibil noted that Amazon’s “flywheel” – from retail and Prime Video to its healthcare acquisitions – is increasingly powered by Amazon Web Services (AWS), which provides the AI cloud infrastructure behind many services binaryverseai.com. AWS now commands about 30% of global cloud market share, and its profits (along with booming advertising revenue) help fund Amazon’s AI “moonshots” binaryverseai.com. CEO Andy Jassy has compared AI’s transformative impact to electricity in its ubiquity binaryverseai.com. In practice, Amazon has embedded AI throughout its operations, from warehouse automation to Alexa voice assistants, and is investing heavily in generative AI services for AWS clients. The message from Amazon’s recent earnings and moves is clear: AI is not a side project for the tech giant but a core part of its long-term strategy – even if it’s less publicized than some competitors’ efforts.

Salesforce and Perplexity Roll Out New AI Tools: Enterprise software leader Salesforce this week launched Agentforce 3, an upgrade to its AI-driven customer support platform. The new system turns chatbots (“agents”) into true virtual teammates for human service reps, complete with a command center that offers live monitoring, session replays, and an Agent Exchange marketplace of over 100 pre-built automations binaryverseai.com. These enhancements have real business impact – Salesforce reported a 233% increase in adoption of its AI agents over six months as companies find that AI can now resolve the bulk of support tickets and dramatically cut handling times binaryverseai.com. Meanwhile, AI startup Perplexity – known for its AI search assistant – unveiled a suite of features transforming its product from simple Q&A into a research and productivity studio. The updated Perplexity Labs can generate reports, slideshows, or even simple web apps from natural language prompts binaryverseai.com. New voice interaction lets users ask questions aloud and get spoken answers, and a file upload feature enables semantic search through documents or meeting transcripts binaryverseai.com. With generous free tiers and privacy options (users can wipe conversation logs), Perplexity’s tool now blends capabilities reminiscent of Notion, ChatGPT and Wolfram Alpha into one AI assistant binaryverseai.com. The trend is clear: companies large and small are rapidly integrating AI into products to boost productivity and offer new capabilities, signaling a competitive advantage for those who harness these tools effectively.

Breakthrough AI Technologies and Research

DeepMind’s AlphaGenome Advances Genomics: Google’s AI research arm DeepMind announced a major breakthrough in AI for genetics. Their new model AlphaGenome can analyze up to 1 million DNA base pairs at a time and predict how genetic mutations will affect gene regulation and function deepmind.google deepmind.google. This unified DNA-sequence model leverages convolutional neural networks and transformers to capture both local DNA motifs and far-ranging gene interactions – it can, for example, detect an enhancer region nearly 980,000 base pairs away and still determine its influence on a target gene binaryverseai.com. In benchmarking tests, AlphaGenome outperformed all prior methods on 22 of 24 key genomics tasks, and even rediscovered a known leukemia-causing mutation that human scientists had identified only after years of research binaryverseai.com. DeepMind has made AlphaGenome available via API for non-commercial research, hoping it will accelerate discoveries in genome science deepmind.google deepmind.google. Researchers are heralding this as a “genomic thunderclap” – effectively making the genome “searchable” with AI binaryverseai.com. By rapidly scoring the effects of DNA variations, AI like AlphaGenome could help uncover genetic drivers of diseases and guide the development of new treatments.

Alibaba Unveils Multimodal Qwen-VLo Model: China’s tech giant Alibaba announced a significant AI milestone with Qwen-VLo, a next-generation multimodal AI model. Building on Alibaba’s Qwen series, Qwen-VLo is a unified vision-language model capable of both understanding images and generating wholly new images from text prompts qwenlm.github.io. “This newly upgraded model not only ‘understands’ the world but also generates high-quality recreations based on that understanding,” the Alibaba Qwen research team explained qwenlm.github.io. In practice, users can feed Qwen-VLo an input image and ask for complex edits, or just describe an image in natural language and have the model create it. Demos showed Qwen-VLo accurately performing tasks like style transfer (e.g. “make this photo look like a Van Gogh painting”), object insertion (“put a red hat on the cat”), and even combined instructions in one go qwenlm.github.io qwenlm.github.io. Uniquely, Qwen-VLo supports open-ended image editing instructions and is multilingual (handling both Chinese and English prompts seamlessly) qwenlm.github.io qwenlm.github.io. The model generates images through a progressive process – refining details from coarse to fine – which leads to coherent and realistic outputs. Alibaba has integrated Qwen-VLo into its Qwen Chat interface as a preview, signaling a push to offer AI that can both see and create, analogous to OpenAI’s Vision-enabled GPT-4. This reflects a broader industry trend toward multimodal AI that can cross between text, vision, and other domains in a unified system.

Self-Improving “SEAL” AI Models: In academic AI research, scientists at MIT introduced an approach for Self-Adapting Language Models (nicknamed “SEAL”) that can learn autonomously from their own outputs. Instead of remaining static after training, a SEAL model can generate practice problems for itself, attempt to solve them, evaluate its answers, and then update its knowledge – all without human intervention. In puzzle-solving benchmarks, a prototype SEAL system boosted its success rate from 0 to 72% through iterative self-training binaryverseai.com. The model uses reinforcement learning to reward itself for improvements and can integrate new data on the fly, though researchers warn it risks “catastrophic forgetting” of older knowledge binaryverseai.com. The promise of SEAL is an AI that “grows like an apprentice” rather than a fixed expert binaryverseai.com. Envision a coding assistant that overnight teaches itself new test cases based on yesterday’s errors, or an educational tutor that refines its lessons after each student interaction – those are the kinds of applications the SEAL concept hints at. While still experimental, the work shows that autonomous improvement is now a design philosophy for AI systems, bringing us a step closer to AI that can continually adapt and refine itself binaryverseai.com.

AI Beats the Lie Detector (Mostly): A new meta-study has shaken up the field of deception detection. Researchers reviewed 98 studies and found that AI-driven systems using convolutional neural networks (CNNs) to analyze human cues can outperform traditional polygraph tests in detecting lies binaryverseai.com. These AI systems process a person’s micro-expressions, eye blinks, vocal tremors, body heat patterns, and even EEG brainwave data to discern truthful vs. deceptive behavior binaryverseai.com. Humans often miss fleeting facial cues like a split-second eyebrow twitch, but machines can catch them at 240 frames per second binaryverseai.com. However, the study also highlighted big caveats: deception signals vary across cultures and genders – a raised brow might indicate doubt in one culture but respect in another binaryverseai.com. Current lie-detection models tend to overfit to regional data, reducing their reliability globally. The review calls for more diverse training data and emphasizes ethical guardrails binaryverseai.com. In short, AI lie detectors are getting better than the old polygraph, but they’re not infallible – context matters. Experts stress that any use of such tools must account for privacy and the risk of false positives, echoing broader debates about AI’s role in surveillance and law enforcement.

AI in Healthcare and Life Sciences

Deep Learning Predicts Postpartum Hemorrhage: In a promising medical AI development, a Chinese research team led by Dr. Wenzhe Zhang reported an AI model that can predict postpartum hemorrhage (PPH) – a leading cause of maternal mortality – before childbirth. By analyzing pregnant women’s MRI scans with a deep-learning “late fusion” model (combining 2D and 3D convolutional neural nets plus radiomics and clinical data), their approach identified high-risk cases with remarkable accuracy. In trials on 581 patients, the AI achieved about 92% sensitivity and 91% specificity in predicting which women would suffer severe bleeding, outperforming other methods auntminnie.com binaryverseai.com. “Earlier identification of the patients at risk of postpartum hemorrhage is critical for optimizing the delivery plan, preparing necessary blood products, and minimizing adverse outcomes,” the researchers noted in Academic Radiology auntminnie.com. With PPH accounting for roughly 25% of maternal deaths worldwide auntminnie.com, such an AI tool could be a lifesaver – allowing doctors to mobilize blood transfusions and surgical teams in advance for those flagged at risk. While further validation is needed before clinical adoption, this study underscores how AI, paired with routine MRI scans, can catch subtle warning signs that human eyes might miss, potentially saving mothers’ lives in childbirth.

AI Microscope Finds “Invisible” Sperm Cells: Another medical breakthrough came from the fertility field. An AI-assisted microscope has demonstrated it can detect viable sperm cells in extremely low-count male infertility cases that stump conventional methods. In one dramatic example, at a fertility clinic a sample was declared hopeless after technicians spent 48 hours searching slides and found zero sperm. The AI system, using a microfluidic chip and computer vision, scanned the sample and flagged 44 viable sperm in under an hour binaryverseai.com. That was enough for a specialized IVF procedure (ICSI – intracytoplasmic sperm injection) and eventually led to a successful pregnancy binaryverseai.com. Crucially, the AI approach avoids the need for toxic dyes or invasive biopsies to find sperm. Experts say this technology could scale up to help many cases of male infertility – for instance by ranking sperm health to pick the best cells, or by extending the technique to evaluate eggs and embryos. In short, what used to be “finding a needle in a haystack” – searching for a few good sperm among millions of cells – can now be done reliably with AI in a fraction of the time. For couples struggling to conceive, such advances mean new hope. It’s a powerful illustration of how AI in the lab is directly changing lives, turning once-impossible scenarios into success stories.

Medical Imaging AI Guards Mothers’ Health: Beyond postpartum hemorrhage, AI is tackling other obstetric risks. Researchers in California and China combined MRI “radiomic” analysis with machine learning to predict placenta accreta spectrum (a dangerous condition where the placenta attaches too deeply) and associated hemorrhage. The ensemble model fusing imaging and clinical data not only predicted hemorrhage events but did so early enough to inform delivery planning binaryverseai.com. In plain terms, this means radiologists with an AI assist can flag high-risk pregnancies weeks before labor. Hospitals can then ensure blood banks are ready and specialists on standby, drastically improving outcomes. This achievement ties into a larger trend: AI-enhanced diagnostics in medical imaging. From detecting cancer on mammograms to assessing brain scans for stroke, AI systems are increasingly acting as a second set of eyes for doctors. In the maternity ward, that extra foresight can be the difference between life and death, especially in regions where maternal care resources are limited.

AI in Robotics and Autonomous Systems

Google DeepMind’s On-Device Robot Brain: One of the most exciting launches came from Google DeepMind’s robotics division, which introduced Gemini Robotics On-Device, a new AI foundation model that runs entirely locally on robots – with no cloud connection needed pymnts.com pymnts.com. This vision-language-action model allows a humanoid robot to perceive its environment and perform complex tasks with low latency and without relying on an internet link. “Since the model operates independent of a data network, it’s helpful for latency-sensitive applications and ensures robustness in environments with intermittent or zero connectivity,” said Carolina Parada, DeepMind’s Head of Robotics pymnts.com pymnts.com. Building on an earlier “Gemini” model revealed in March, the On-Device version is designed for bimanual (two-armed) robots and can quickly learn new tasks through fine-tuning. Google reports that the system can handle everyday actions like unzipping bags, folding laundry, pouring liquids, and even drawing a card from a deck pymnts.com. Developers showed that with only 50–100 demonstrations, the model can generalize its skills to a new task, reflecting a huge leap in robot dexterity and adaptability pymnts.com. This is also Google DeepMind’s first major robotics model that developers can fine-tune themselves pymnts.com, opening the door for customization. The significance of Gemini On-Device is that robots can now “think” and react in real time on the edge – crucial for industries like manufacturing or home robotics where split-second decision-making and privacy (keeping data on the device) are paramount. As one tech outlet quipped, with this advance “the robot now thinks locally and acts instantly” binaryverseai.com, which could accelerate the arrival of helpful humanoid robots in the real world.

ABB’s Heavy-Lifting Warehouse Robot: In the industrial robotics arena, ABB unveiled the Flexley Mover P603, an autonomous mobile robot with a deceptively small form factor. Roughly the size of a coffee table, this squat vehicle can carry loads up to 1,500 kg (1.5 tons) binaryverseai.com – an impressive feat for its footprint. The P603 navigates using visual SLAM (simultaneous localization and mapping), meaning it can map a warehouse floor on the fly without needing special QR codes or tracks binaryverseai.com. It also features an active suspension to handle rough floors and can position heavy pallets with 5 mm precision while moving at 2 m/s binaryverseai.com. Perhaps most appealing to factory managers, the robot’s workflow can be configured via a drag-and-drop interface in ABB’s software studio, rather than complex programming binaryverseai.com. In other words, setting up the robot’s routes and tasks is almost as easy as building a playlist. The P603 arrives at a time when factories and warehouses are increasingly aiming for flexible automation – replacing fixed conveyor belts and guided vehicles with free-roaming robots that can be re-tasked on the fly. ABB’s offering, noted in an industry roundup this week, is “another brick” in the wall of AI-driven automation sweeping logistics binaryverseai.com. As supply chains adapt to rapid e-commerce growth and labor shortages, such intelligent robots are becoming indispensable.

A prototype of a mosquito-sized surveillance drone, as unveiled by a Chinese defense lab tomshardware.com. State media footage showed the tiny bionic drone – as small as an insect – being held between two fingers.

China’s Mosquito-Sized Spy Drone: It sounds like science fiction, but Chinese researchers have built a drone the size of an actual mosquito. This week, China’s state TV CCTV-7 aired footage of the tiny robotic flyer, which a student from the National University of Defense Technology demonstrated by pinching it between his fingertips tomshardware.com. The mosquito drone comes in at least two variants – one with two wings and one with four – and is designed for covert surveillance missions tomshardware.com. While technical specs remain secret (it’s unclear what sensors or battery life it has, given the insect-scale hardware), experts say the mere reveal of this project signals China’s intent to push micro-UAV technology to new extremes binaryverseai.com. Such miniaturized drones could potentially slip into buildings or hover undetected in urban environments where larger drones can’t go, raising complex ethical and security questions. Defense analysts note that many nations are working on insect-scale drones for reconnaissance; the challenges include achieving useful range and transmitting data back reliably given the tiny power supply tomshardware.com. China’s prototype is likely still in research stages (no evidence yet it’s deployed in the field tomshardware.com), but it shows how far drone innovation has come – quite literally bringing surveillance down to bug size. The development has prompted discussions about countermeasures and privacy, as society grapples with the idea that a mosquito buzzing by your ear might not be a mosquito at all.

AI Policy, Ethics, and Expert Perspectives

Landmark Copyright Ruling on AI Training Data: A U.S. federal judge issued a much-anticipated ruling with huge implications for AI companies and copyright law. In a lawsuit against AI startup Anthropic (maker of the Claude chatbot), Judge William Alsup held that using copyrighted books to train an AI can qualify as fair use under U.S. law – a major win for the AI industry apnews.com. Alsup’s decision likened an AI training on thousands of books to a human writer reading classics like Dickens to inspire new work, calling the AI’s output “quintessentially transformative” and not a mere copy apnews.com. However, the judge drew a critical line: while the analysis (training) may be fair use, the method of acquiring the data still matters. In Anthropic’s case, the firm had obtained many books from “shadow libraries” of pirated e-books, essentially illegal downloads apnews.com apnews.com. Judge Alsup ruled Anthropic must face trial for copyright theft because “Anthropic had no entitlement to use pirated copies for its library” even if the end use was transformative apnews.com. In effect, the court dismissed the claim that training itself was infringement, but left the door open to liability if the training data was obtained illicitly. This split decision sets a precedent as perhaps the first judicial take on AI training and fair use. It suggests that “permissionless learning survives” – AI companies can learn from copyrighted material without direct licenses – but data sourcing shortcuts don’t binaryverseai.com. Going forward, AI developers are on notice to clean up their training pipelines: scraping the internet or pirate sites could incur legal risk, while using legitimately purchased or public-domain data will be the safer route binaryverseai.com. The ruling comes as similar copyright lawsuits pile up against OpenAI and others apnews.com, and it will likely influence how those cases proceed. Anthropic, for its part, said it was pleased the judge recognized AI training as transformative and in line with copyright’s purpose of fostering new creativity apnews.com. The trial on the remaining issues is set for December, and the AI community will be watching closely as legal frameworks catch up with technological reality.

AI’s Energy Appetite Under Scrutiny: As AI models grow ever larger, concerns are mounting about their environmental and energy impact. This week tech columnist Joanna Stern took a deep dive into the question “How much energy does your AI prompt use?” – and the findings are eye-opening. Even seemingly trivial AI tasks can consume significant power. For instance, generating a single 6-second AI video clip can use “anywhere between 20 and 110 watt-hours” of energy livemint.com. At the high end, that’s roughly the electricity needed to run an electric grill for 10 minutes, which Stern demonstrated by cooking a steak with the equivalent energy an AI video request might burn livemint.com. In practical terms, two short AI-generated videos might gobble the same power as grilling an entire dinner livemint.com. And larger AI workloads scale up dramatically: training large language models involves thousands of such GPU-heavy tasks, drawing on megawatt-hours of electricity and large volumes of water for cooling data centers livemint.com livemint.com. The mysterious journey of an AI prompt – from a user’s laptop to a far-off GPU server and back – is often hidden from consumers, but Stern’s report (and a growing body of research) is pulling back the curtain on this “energy drain” linkedin.com linkedin.com. Researchers like Sasha Luccioni at Hugging Face have even started an AI Energy consumption leaderboard, benchmarking different models’ power use livemint.com. The good news is that hardware is improving: Nvidia’s latest AI chips are reportedly 30× more energy-efficient than those of just a year ago, according to the company’s sustainability lead livemint.com. Tech firms also tout efforts to shift to cleaner energy sources for their data centers livemint.com. But efficiency gains may be outpaced by sheer growth in AI usage – more models, more users, more queries mean more energy overall, even if each operation gets a bit greener livemint.com livemint.com. Stern and others suggest transparency is key: if users saw an “energy cost” readout for each AI query, they might think twice about frivolous uses linkedin.com. Ultimately, the industry faces a dual challenge of curbing AI’s carbon footprint while still innovating. The takeaway for now: AI isn’t magic – it runs on electricity, and lots of it. As one executive quipped, AI is only as sustainable as the power (and water) we feed it livemint.com, so future breakthroughs must include not just smarter models, but more energy-savvy ones too.

Experts Debate AI’s Unpredictable Trajectory: The rapid progress of AI has even its pioneers sounding both optimistic and cautionary notes. Ilya Sutskever, the co-founder and chief scientist of OpenAI, made waves with a public warning that AI’s evolution could spiral in unforeseen ways. “AI is going to be both extremely unpredictable and unimaginable,” Sutskever said in a recent interview, cautioning that advanced AI systems might one day start improving themselves without human oversight analyticsindiamag.com. He suggested this could trigger “rapid and uncontrollable progress,” making it hard for humans to understand or manage what comes next analyticsindiamag.com. This stark warning came alongside Sutskever’s reflections on the concept of an “intelligence explosion” – the idea that a sufficiently advanced AI could keep rewriting a better version of itself, leading to exponential gains in capability. On the positive side, Sutskever noted such AI could yield “incredible healthcare” breakthroughs, curing diseases and extending human lifespan analyticsindiamag.com. Yet he paired that optimism with concern about how we would respond if AI became that powerful analyticsindiamag.com. His comments underscore a broader discussion in the AI community: how to balance the promises of AI (in medicine, science, etc.) with the perils of losing control or oversight. Notably, Sutskever recently left OpenAI to found a new venture, Safe Superintelligence, aimed at ensuring future AI remains beneficial analyticsindiamag.com analyticsindiamag.com. His stance resonates with calls from other tech leaders for robust AI safety research now, not later. The fact that one of AI’s leading architects openly worries about scenarios that read like science fiction – self-evolving AI, beyond human understanding – shows that the ethical and existential questions around AI are no longer academic. They’re here and need addressing through global collaboration, thoughtful regulation, and continued research into AI alignment with human values.

AI and the Future of Work – A Labor Gap Warning: Amid the focus on high-tech advances, a stark reminder came from the manufacturing world: Who will build the AI future? Ford CEO Jim Farley, speaking at the Aspen Ideas Festival, warned that while AI-driven productivity is booming for white-collar roles, the supply of skilled blue-collar trades is drying up binaryverseai.com. Farley noted that factories still rely on skilled electricians, welders, and technicians – jobs that AI and robots have not yet filled beyond perhaps 10-20% of tasks binaryverseai.com. He gave a vivid example: at an auto plant, a German line worker once ingeniously fixed a stuck tailgate with a bicycle tire – a creative, on-the-spot hack no algorithm would have predicted binaryverseai.com. That kind of human improvisation remains critical on factory floors. But younger generations are less often entering trades, and as current tradespeople retire, there’s concern that industries could hit a bottleneck: you can’t scale up EV factories or infrastructure projects without enough human hands to do the physical work. Farley advocated for investing in trade education and reframing these careers as vital high-tech jobs of the future (which, in a sense, they are – today’s electricians often work alongside automation and advanced machinery) binaryverseai.com. He even cast it as a national security issue (“domestic manufacturing is our defense”), implying the competitiveness of nations depends on having people who can build the innovations that AI dreams up binaryverseai.com. The takeaway is a nuanced one: AI will change jobs, but it can also create new demands on the workforce. As AI handles routine cognitive tasks and robots take over repetitive manual ones, the remaining jobs will require more skill, adaptability and often interdisciplinary know-how (combining, say, carpentry with programming for a smart home installer). Policymakers and companies are thus urged to plan for workforce development, so that society isn’t caught with a million prompt engineers but too few plumbers. In Farley’s blunt terms, “America needs a blueprint” to ensure that technological progress and human labor progress hand-in-hand binaryverseai.com.

Sources: The information above is drawn from a range of authoritative sources, including company announcements, expert interviews, and news outlets covering AI. Key references include an Associated Press report on the Anthropic copyright ruling apnews.com apnews.com, analysis by The Wall Street Journal on AI energy consumption livemint.com livemint.com, statements from AI leaders like Ilya Sutskever via Analytics India Magazine analyticsindiamag.com analyticsindiamag.com, and corporate news from outlets such as Cointelegraph (Meta’s hires and deals) cointelegraph.com cointelegraph.com. Cutting-edge research findings were sourced from publications like DeepMind’s official blog (AlphaGenome) deepmind.google, Academic Radiology via AuntMinnie (AI for postpartum hemorrhage) auntminnie.com auntminnie.com, and tech news sites (e.g. Tom’s Hardware for the mosquito drone) tomshardware.com. These developments collectively paint a picture of an AI landscape that is rapidly evolving – with expanding capabilities and influence – even as society grapples with understanding and guiding this transformation. Each week’s progress brings both excitement and introspection about the role of AI in our lives, economy, and future.