Latest Tech News

Stay updated with the latest in technology, AI, cybersecurity, and more

Filtered by: model Clear Filter

Hierarchical Modeling (H-Nets)

The best AI architectures in use today treat all inputs equally. They process each input with the same amount of compute, without explicitly grouping related inputs into higher level concepts. While these architectures have achieved impressive results across domains, this lack of hierarchy has some fundamental limitations. Models have difficulty learning from high resolution, raw data, requiring inputs to be pre-processed into meaningful tokens for strong performance. The use of hand-crafted p

Finally, a dev kit for designing on-device, mobile AI apps is here: Liquid AI’s LEAP

Want smarter insights in your inbox? Sign up for our weekly newsletters to get only what matters to enterprise AI, data, and security leaders. Subscribe Now Liquid AI, the startup formed by former Massachusetts Institute of Technology (MIT) researchers to develop novel AI model architectures beyond the widely-used “Transformers“, today announced the release of LEAP aka the “Liquid Edge AI Platform,” a cross-platform software development kit (SDK) designed to make it easier for developers to int

Tesla Arrives in India with Ridiculous Prices

In a nation where the average salary is around $4,200 a year, the all-electric carmaker is launching a $70,000 vehicle. The question is why. The wait is over. Tesla has finally launched in India. The all-electric carmaker has officially opened up sales in the world’s third-largest auto market, hoping that a foothold in this fast-growing economy could help counteract the company’s recent global slump. On Tesla’s newly launched India webpage, customers can now place orders. But the offering is

I replaced my Galaxy S25 Plus with the S25 Edge (and Samsung may do the same)

Kerry Wan/ZDNET The Galaxy S25 Edge was my unexpected favorite phone in Samsung's S-series lineup this year. I previously wrote about how it had one major fault, but is the compact phone for content consumption in 2025. It just might not make sense to exist alongside the Plus model with the adoption of silicon-carbon batteries -- and Samsung might agree. Also: Samsung's ultra-thin Galaxy phone is a greatly misunderstood device - and I shamelessly love it According to a new report from The Ele

Tesla Arrives in India With Cars Priced for the 1%

The wait is over. Tesla has finally launched in India. The all-electric carmaker has officially opened up sales in the world’s third-largest auto market, hoping that a foothold in this fast-growing economy could help counteract the company’s recent global slump. On Tesla’s newly launched India webpage, customers can now place orders. But the offering is limited: Tesla is only selling the Model Y SUV, available in two variants—Rear-Wheel Drive with a base price of ₹59,89,000 ($69,647) and Long-

Research leaders urge tech industry to monitor AI’s ‘thoughts’

AI researchers from OpenAI, Google DeepMind, Anthropic, as well as a broad coalition of companies and nonprofit groups, are calling for deeper investigation into techniques for monitoring the so-called thoughts of AI reasoning models in a position paper published Tuesday. A key feature of AI reasoning models, such as OpenAI’s o3 and DeepSeek’s R1, are their chains-of-thought or CoTs — an externalized process in which AI models work through problems, similar to how humans use a scratch pad to wo

The Hidden Costs of AI: How Small Inefficiencies Stack Up

AI is having a major impact on society, from consumer technologies to driving businesses. But among all these large language models and deep neural networks, there are lurking inefficiencies that most people aren’t taking into account. Wasted computational power, hidden costs, the environmental footprint, and more. Whether these inefficiencies are big or small, they all add up. With 58% of companies planning on increasing their investments in AI this year, it’s essential these decision-makers k

Mistral releases Voxtral, its first open source AI audio model

As AI systems become more capable, speech is fast becoming the default way we communicate with machines. French AI startup Mistral has jumped into the audio race with its first open model, aiming to challenge the dominance of walled-off corporate systems with open-weight alternatives. On Tuesday, Mistral announced the release of Voxtral, its first family of audio models aimed at businesses. The company is pitching Voxtral as the first open model that’s capable of deploying “truly usable speech

Show HN: We made our own inference engine for Apple Silicon

uzu A high-performance inference engine for AI models on Apple Silicon. Key features: Simple, high-level API Hybrid architecture, where layers can be computed as GPU kernels or via MPSGraph (a low-level API beneath CoreML with ANE access) Unified model configurations, making it easy to add support for new models Traceable computations to ensure correctness against the source-of-truth implementation Utilizes unified memory on Apple devices Quick Start First, add the uzu dependency to your

Topics: let model run session uzu

xAI says it has fixed Grok 4’s problematic responses

When xAI launched Grok 4 last week, the company claimed the large language model outperformed several competitors on different benchmarks. But the Grok account on X that runs off the model immediately showed there were some major issues: it started saying its surname was “Hitler”, tweeted antisemitic messages, and seemed to reference Elon Musk’s posts when asked about controversial topics, siding with the xAI owner’s views as a result. xAI soon afterwards apologized for Grok’s behavior. On Tue

AI text-to-speech programs could “unlearn” how to imitate certain people

AI companies generally keep a tight grip on their models to discourage misuse. For example, if you ask ChatGPT to give you someone’s phone number or instructions for doing something illegal, it will likely just tell you it cannot help. However, as many examples over time have shown, clever prompt engineering or model fine-tuning can sometimes get these models to say things they otherwise wouldn’t. The unwanted information may still be hiding somewhere inside the model so that it can be accessed

Exclusive: Google Pixel Watch 4 to use the same processor as Pixel Watch 2 and 3

TL;DR Google’s next-gen Pixel Watch 4 is keeping the same Snapdragon W5 Gen 1 chip that’s been used since the Pixel Watch 2. Both 41mm and 45mm models should receive slightly larger batteries. As we approach Google’s 2025 Pixel event, we already know almost everything about its next generation of phones, but the same can’t be said about the upcoming Pixel Watch 4. Other than some renders and a color options leak, there’s been nothing concrete said about the new smartwatches. We heard some info

Meta built its AI reputation on openness — that may be changing

Top members of Meta’s new Superintelligence Lab discussed pivoting away from the company’s powerful open source AI model, Behemoth, and instead developing a closed model, reports The New York Times. Sources told The Times that Meta had completed training on Behemoth, but delayed its release due to underwhelming internal performance. When the new Superintelligence Lab launched, testing on the model reportedly halted. The discussions are just that — discussions. Meta CEO Mark Zuckerberg would st

Context Rot: How increasing input tokens impacts LLM performance

Recent developments in LLMs show a trend toward longer context windows, with the input token count of the latest models reaching the millions. Because these models achieve near-perfect scores on widely adopted benchmarks like Needle in a Haystack (NIAH) [1], it’s often assumed that their performance is uniform across long-context tasks. However, NIAH is fundamentally a simple retrieval task, in which a known sentence (the “needle”) is placed in a long document of unrelated text (the “haystack”)

OpenAI Hits Pause on Its Meta Killer

OpenAI is delaying the release of its much-anticipated open-weight AI model, citing the need for “additional safety tests” and last-minute concerns over “high-risk areas,” CEO Sam Altman announced on X (formerly Twitter). The decision lands in the middle of a brutal AI arms race, particularly with Meta, which has been aggressively poaching OpenAI talent and championing open-source models like Llama 3. The model, which was slated to drop this week, would be OpenAI’s first major open-weight syste

AI slows down open source developers. Peter Naur can teach us why

AI slows down open source developers. Peter Naur can teach us why. AI slows down open source developers. Peter Naur can teach us why. Metr recently published a paper about the impact AI tools have on open-source developer productivity1. They show that when open source developers working in codebases that they are deeply familiar with use AI tools to complete a task, then they take longer to complete that task compared to other tasks where they are barred from using AI tools. Interestingly the

xAI starts offering Grok to US government agencies

Just days after apologizing for Grok's recent hard turn toward antisemitism, xAI has announced a suite of AI products for government use. Grok for Government brings together the company's latest commercial products, including Grok 4 and Deep Search, with special considerations given to the needs of federal, state and local agencies. To that end, xAI says it will design custom models for specific national security and research customers. It will also develop specialized AI applications for use i

New Grok AI model surprises experts by checking Elon Musk’s views before answering

An AI model launched last week appears to have shipped with an unexpected occasional behavior: checking what its owner thinks first. On Friday, independent AI researcher Simon Willison documented that xAI's new Grok 4 model searches for Elon Musk's opinions on X (formerly Twitter) when asked about controversial topics. The discovery comes just days after xAI launched Grok 4 amid controversy over an earlier version of the chatbot generating antisemitic outputs, including labeling itself as "Mech

A Clear Sign Just Emerged That Tesla Is Panicking

Tesla sales are so abysmal in Canada that Elon Musk has made the drastic choice to dramatically lower the price of the new Model Y as it attempts to climb out of the hole. As Electrek reports, a Model Y in Canada now costs $65,ooo CAD ($47,465 USD), which is a cool $20,000 CAD cheaper than it was just a few days prior. We'll get into the hows and whys in a second, but first some backstory. Earlier this year, after Trump raised the specter of tariffs against its northern neighbor, Canada moved

As Sales Drop, Tesla Makes a Big Gamble on India

Tesla is finally entering India. The all-electric carmaker will open its first showroom in the country on July 15 in Mumbai, marking a long-awaited entry into one of the largest and fastest-growing automotive markets in the world. Invitations to the launch event began circulating on social media late last week. “Launch Event. Exclusive Invite,” the sleek black cards read. “Tesla Experience Center BKC.” The event is scheduled to take place from 11:00 a.m. to 12:30 p.m. local time, with guests as

Tesla Makes a Desperate Move in Canada as Sales Collapse

Tesla is making a desperate move to stay alive in Canada as its sales suffer from the fallout of Donald Trump’s trade war. In a quiet but shocking update to its website, the electric vehicle maker has drastically lowered the price of its best-selling Model Y SUV by a staggering $20,000. The move is a direct response to a brutal trade dispute that has crippled Tesla’s Canadian operations. In retaliation for tariffs imposed by the Trump administration, Ottawa slapped a 25% surtax on all cars impo

Alibaba-backed Moonshot releases new Kimi AI model that beats ChatGPT, Claude in coding — and it costs less

An AI sign at the MWC Shanghai tech show on June 19, 2025. Bloomberg | Bloomberg | Getty Images BEIJING — The latest Chinese generative artificial intelligence model to take on OpenAI's ChatGPT is offering coding capabilities — at a lower price. Alibaba-backed startup Moonshot released on late Friday night its Kimi K2 model: a low-cost, open source large language model — the two factors that underpinned China-based DeepSeek's industry disruption in January. Open-source technology provides sourc

Topics: ai k2 kimi model openai

Show HN: ArchGW – An intelligent edge and service proxy for agents

Arch is a proxy server designed as a modular edge and AI gateway for agentic apps Arch handles the pesky low-level work in building agentic apps — like applying guardrails, clarifying vague user input, routing prompts to the right agent, and unifying access to any LLM. It’s a language and framework friendly infrastructure layer designed to help you build and ship agentic apps faster. Quickstart • Demos • Build agentic apps with Arch • Route LLMs • Documentation • Contact Overview AI demos a

Tesla Makes a Desperate Move in Canada As Sales Collapse

Tesla is making a desperate move to stay alive in Canada as its sales suffer from the fallout of Donald Trump’s trade war. In a quiet but shocking update to its website, the electric vehicle maker has drastically lowered the price of its best-selling Model Y SUV by a staggering $20,000. The move is a direct response to a brutal trade dispute that has crippled Tesla’s Canadian operations. In retaliation for tariffs imposed by the Trump administration, Ottawa slapped a 25% surtax on all cars impo

How to scale RL to 10^26 FLOPs

TLDR: Reinforcement learning (RL) is the next training technique for building frontier-level AI models. To make it better, we need to train on more data. The current approach of scaling many environments simultaneously is messy and complicated. Instead, I propose we find a way to do next-token prediction on the Web using RL. This way, we learn to reason from general web data, instead of just math and code. I’ve spent a good part of the past year in denial. I was in denial because when OpenAI r

Study warns of ‘significant risks’ in using AI therapy chatbots

Therapy chatbots powered by large language models may stigmatize users with mental health conditions and otherwise respond inappropriately or even dangerously, according to researchers at Stanford University. While recent coverage in The New York Times and elsewhere has highlighted the role that ChatGPT may play in reinforcing delusional or conspiratorial thinking, a new paper titled “Expressing stigma and inappropriate responses prevents LLMs from safely replacing mental health providers” exam

A new Martian climate model suggest a mostly cold, harsh environment

The Curiosity rover was sent up the Mount Sharp, the biggest sediments stack on Mars. On the way, it collected samples that indicated a portion of carbon dioxide in the Martian atmosphere might have been sequestered in the sedimentary rocks, just as it happens with limestone on Earth. This would have drawn carbon dioxide out of the atmosphere, reducing the greenhouse effect that warmed the planet. Based on these findings, a team of scientists led by Benjamin Tutolo, a researcher at the Universi

New Video-Generating AI Trained 100 Percent on Public Domain Films

Few tech products have been as broadly contentious as video-generating artificial intelligence. These complex algorithms, which cleave millions of datapoints together into seconds-long gobs of video, are notoriously trained on proprietary material, leading to widespread ethical and legal concerns. (That's before we even mention how much energy it takes to synthesize an AI video.) Tech billionaires tend to argue that this is simply the way things need to be — if you want AI, we need to feed it

Kimi k2 largest open source SOTA model?

📰 Tech Blog | 📄 Paper Link (coming soon) 1. Model Introduction Kimi K2 is a state-of-the-art mixture-of-experts (MoE) language model with 32 billion activated parameters and 1 trillion total parameters. Trained with the Muon optimizer, Kimi K2 achieves exceptional performance across frontier knowledge, reasoning, and coding tasks while being meticulously optimized for agentic capabilities. Key Features Large-Scale Training: Pre-trained a 1T parameter MoE model on 15.5T tokens with zero train

Topics: acc kimi model shot tool