Tag: Large language models

Exclusive: Complex Chaos thinks AI can help people find common ground | TechCrunch

[ad_1]

Making democracy work isn’t easy, as recent events have made clear. Some critics would argue that technology is making it worse. But one startup is hoping that AI could help bridge some differences instead of widen them.

“I had an a-ha moment one day when I realized people are asking AI to explain something like they’re five years old,” Tommy Lorsch, co-founder and CEO of Complex Chaos, told TechCrunch. “What if we use it as a facilitator to help people understand each other and find common ground?”

He and co-founder Maya Ben Dror are developing tools to help people arrive at a consensus. One of their first test cases involved climate negotiations, but it really doesn’t matter what the issue is. His goal is to foster cooperation and shorten the time it takes for groups to come to agreement.

“Everyone is building software for collaboration like Slack, Google Docs, whatever,” Lorsch said. “Cooperation is a different piece.”

Facilitating cooperation isn’t something that scales well, he said. Typically, trained facilitators will spend time with groups to help them arrive at a consensus, but that process can slow down when negotiations or preparations happen across time zones or even in different rooms.

Lorsch was buoyed by a recent LLM developed by Google called the Habermas Machine, which was developed explicitly with that goal in mind. “This is basically an AI that generates group consensus statements where people feel represented both majority and minority point of view,” he said.

Lorsch and Ben Dror recently trialed their startup’s tool to help young delegates from nine African nations prepare for climate-related negotiations at a United Nations campus in Bonn, Germany. The tool incorporates both Google’s Habermas Machine and OpenAI’s ChatGPT to generate questions, come up with goals for conversations, and help summarize long documents.

Techcrunch event

San Francisco
|
October 27-29, 2025

The goal, Ben Dror said, is to help the delegates arrive at consensus as a bloc before they began negotiations with others.

Ideally, the tool would help speed things during negotiations, too. When blocs, or delegates fro groups of aligned countries, encounter new information in the process of a large negotiating session, they often need to regroup to process the new information. “Blocs are usually the reason why negotiations have to stop. The bloc has to come out, renegotiate, reposition, and then go back in. And that creates a lot of friction,” Ben Dror said. Complex Chaos hopes that its tool can help shorten that time.

In the trial with the delegates from African countries, Complex Chaos said that participants reported up to a 60% reduction in the time it took to coordinate, and that 91% of participants said that the AI tool helped them see perspectives they would have otherwise missed.

Complex Chaos is also pitching its cooperation tool to companies, including tech companies and large consultancies. “Strategic planning by AI and basically the same problem,” Lorsch said. “The annual strategic planning process of most companies takes about three months of the year with back and forth negotiations, multi-layer, across time zones, across teams, and so on and so forth.”

But Lorsch and Ben Dror are most enthusiastic when talking about climate negotiations.

“If AI can shorten these processes, simplify them, then we’d be so much better off. Not just for climate, for anything sustainability, for any big challenge that we’re facing,” Ben Dror said.

[ad_2]

Tim De Chant

Source link

September 29, 2025
How South Korea plans to best OpenAI, Google, others with homegrown AI | TechCrunch

[ad_1]

From tech giants to startups, South Korean players are developing large language models tailored to their own language and culture, ready to compete with global heavyweights like OpenAI and Google.

Last month, the nation launched its most ambitious sovereign AI initiative to date, pledging ₩530 billion, (about $390 million), to five local companies building large-scale foundational models.

The move underscores Seoul’s desire to cut reliance on foreign AI technologies, hoping to strengthen national security and keep a tighter control over data in the AI era.

The organizations picked by the Ministry of Science and ICT to compete were LG AI Research, SK Telecom, Naver Cloud, NC AI, and the startup Upstage.

Every six months, the government will review the first cohort’s progress, cut underperformers, and continue funding the frontrunners until just two remain to lead the country’s sovereign AI drive.

Each player is bringing a different advantage to South Korea’s AI race. TechCrunch spoke with several of the selected companies about how they plan to take on OpenAI, Google, Anthropic and the rest on their home turf. NC AI declined to comment.

LG AI Research: Exaone

LG AI Research, the R&D unit of South Korean giant LG Group, offers Exaone 4.0, a hybrid reasoning AI model. The latest version blends broad language processing with the advanced reasoning features first introduced in the company’s earlier Exaone Deep model.

Techcrunch event

San Francisco
|
October 27-29, 2025

Exaone 4.0 (32B) already scores reasonably well against competitors on Artificial Analysis’s Intelligence Index benchmark (as does Upstage’s Solar Pro2). But it plans to improve and move up the ranks through its deep access to real-world industry data ranging from biotech to advanced materials and manufacturing.

It’s coupling that data with a focus on refining the data before feeding to the models to train. Instead of chasing sheer scale, LG wants to make the entire process more intelligent, so its AI can deliver real, practical value that goes beyond what general-purpose models can offer. “This is our fundamental approach,” co-head Honglak Lee told TechCrunch.

LG is improving its models via familiar tactics: offering them through APIs, then using the real-world data generated by users of those services to train the model to improve.

“As LG’s models improve, our partners can deliver better services, which in turn generate greater economic value and even richer data,” he said.

However, instead of chasing massive GPU clusters, LG AI Research is focusing on efficiency, getting the most out of every chip, and creating industry-specific models, he mentioned. The goal isn’t to outspend the global giants but to outsmart them with high-performing, yet more efficient, AI.

South Korea’s telco giant SK Telecom (SKT) launched its personal AI agent A. (pronounced A-dot) service way back in late 2023 and just rolled out its new large language model, A.X, this July.

Built on top of the Chinese open source model from Alibaba Cloud, Qwen 2.5, A.X 4.0 comes in two models, a hefty 72-billion-parameter version and a lighter 7B version.

SK says that A.X 4.0 processes Korean inputs about 33% more efficiently than GPT-4o did, underscoring its local language edge. (OpenAI’s GPT 5.0 comparison data is not available.) SKT also open sourced its A.X 3.1 models earlier this summer. Meanwhile, the A. service offers features like AI call summaries and auto-generated notes. As of August 2025, it’s already pulled in about 10 million subscribers.

SK’s edge is its versatility, because it has access to information from its telecom network ranging from navigation to taxi-hailing.

“SK Telecom’s role is to act as a bridge between cutting-edge model research and real-world impact. With our telecom infrastructure, extensive user base and proven service like A., we bring AI directly into everyday life, whether in customer service, mobility, or manufacturing,” Taeyoon Kim, head of the foundation model office at SK Telecom, told TechCrunch.

SK Telecom is also investing in AI infrastructure, using GPUaaS, South Korea’s largest GPU-based service, and building a new hyperscale AI data center with AWS. Whatever it lacks, it is partnering to obtain.

“We’re building a full-stack ecosystem with Korean AI chipmaker Rebellions, securing trusted data partnerships through work with the government and universities, and fostering a global research network,” said Kim. “That includes projects like our collaboration with MIT (MGAIC), which applies foundation models to advanced manufacturing and battery and semiconductor innovation.”

Naver Cloud: HyperCLOVA X

Naver Cloud, the cloud services arm of South Korea’s leading internet company, introduced its large language model, HyperClova, in 2021. Two years later, it unveiled an upgraded version, HyperCLOVA X, along with new products powered by the technology: CLOVA X, an AI chatbot, and Cue, a generative AI-driven search engine positioned as a rival to Microsoft’s CoPilot-enhanced Bing and Google’s AI Overview. It also unveiled this year its multimodal reasoning AI model, HyperCLOVE X Think.

Naver Cloud believes the true power of LLMs is to serve as “connectors” linking legacy systems and siloed services to improve usefulness, according to a Naver spokesperson.

Naver stands out as Korea’s only company — and one of the few in the world — that can genuinely claim to have an “AI full stack.” It built its HyperCLOVA X model from scratch and runs the massive data centers, cloud services, AI platforms, applications, and consumer services that bring the technology to life, the spokesperson explained.

Similar to Google — but tuned for South Korea — Naver is embedding its AI into core services like search, shopping, maps, and finance. Its advantage is real-world data. It’s AI Shopping Guide, for instance, offers recommendations based on what people actually want to buy. Other services include CLOVA Studio, which lets businesses build custom generative AI, and CLOVA Carecall, an AI-powered check-in service geared for seniors living alone.

The Naver spokesperson says besting global AI giants like OpenAI and Google hinges on two things: perfecting its “recipe” for models and securing the capital to scale them. Even so, rather than chasing size, the company emphasizes sophistication, arguing its AI is already globally competitive at comparable scales.

Upstage’s Solar Pro 2

Upstage is the only startup competing in the project. Its Solar Pro 2 model, launched last July, was the first Korean model recognized as a frontier model by Artificial Analysis, putting it in the ring with OpenAI, Google, Meta, and Anthropic, according to Soon-il Kwon, executive vice president at Upstage.

While most frontier models have 100 billion to 200 billion parameters, Solar Pro 2 — with just 31 billion — performs better for South Koreans and is more cost-effective, Kwon told TechCrunch.

“Solar Pro 2 has outperformed global models on major Korean benchmarks. With this project, Upstage aims to achieve a Korean language performance of 105% of the global standard,” Kwon said.

Upstage aims to differentiate itself by focusing on real business impact, not just benchmarks, he said. So it is developing specialized models for industries like finance, law, and medicine, while pushing to build a Korean AI ecosystem led by “AI-native” startups.

[ad_2]

Kate Park

Source link

September 27, 2025
Giga ML wants to help companies deploy LLMs offline | TechCrunch

[ad_1]

AI is all the rage — particularly text-generating AI, also known as large language models (think models along the lines of ChatGPT). In one recent survey of ~1,000 enterprise organizations, 67.2% say that they see adopting large language models (LLMs) as a top priority by early 2024.

But barriers stand in the way. According to the same survey, a lack of customization and flexibility, paired with the inability to preserve company knowledge and IP, were — and are — preventing many businesses from deploying LLMs into production.

That got Varun Vummadi and Esha Manideep Dinne thinking: What might a solution to the enterprise LLM adoption challenge look like? In search of one, they founded Giga ML, a startup building a platform that lets companies deploy LLMs on-premise — ostensibly cutting costs and preserving privacy in the process.

“Data privacy and customizing LLMs are some of the biggest challenges faced by enterprises when adopting LLMs to solve problems,” Vummadi told TechCrunch in an email interview. “Giga ML addresses both of these challenges.”

Giga ML offers its own set of LLMs, the “X1 series,” for tasks like generating code and answering common customer questions (e.g. “When can I expect my order to arrive?”). The startup claims the models, built atop Meta’s Llama 2, outperform popular LLMs on certain benchmarks, particularly the MT-Bench test set for dialogs. But it’s tough to say how X1 compares qualitatively; this reporter tried Giga ML’s online demo but ran into technical issues. (The app timed out no matter what prompt I typed.)

Even if Giga ML’s models are superior in some aspects, though, can they really make a splash in the ocean of open source, offline LLMs?

In talking to Vummadi, I got the sense that Giga ML isn’t so much trying to create the best-performing LLMs out there but instead building tools to allow businesses to fine-tune LLMs locally without having to rely on third-party resources and platforms.

“Giga ML’s mission is to help enterprises safely and efficiently deploy LLMs on their own on-premises infrastructure or virtual private cloud,” Vummadi said. “Giga ML simplifies the process of training, fine-tuning and running LLMs by taking care of it through an easy-to-use API, eliminating any associated hassle.”

Vummadi emphasized the privacy advantages of running models offline — advantages likely to be persuasive for some businesses.

Predibase, the low-code AI dev platform, found that less than a quarter of enterprises are comfortable using commercial LLMs because of concerns over sharing sensitive or proprietary data with vendors. Nearly 77% of respondents to the survey said that they either don’t use or don’t plan to use commercial LLMs beyond prototypes in production — citing issues relating to privacy, cost and lack of customization.

“IT managers at the C-suite level find Giga ML’s offerings valuable because of the secure on-premise deployment of LLMs, customizable models tailored to their specific use case and fast inference, which ensures data compliance and maximum efficiency,” Vummadi said.

Giga ML, which has raised ~$3.74 million in VC funding to date from Nexus Venture Partners, Y Combinator, Liquid 2 Ventures, 8vdx and several others, plans in the near term to grow its two-person team and ramp up product R&D. A portion of the capital is going toward supporting Giga ML’s customer base, as well, Vummadi said, which currently includes unnamed “enterprise” companies in finance and healthcare.

[ad_2]

Kyle Wiggers

Source link

December 28, 2023
The Week’s Hottest Takes, From Scott Pilgrim To TLOU 2

[ad_1]

Gamers are a passionate bunch, and we’re no exception. These are the week’s most interesting perspectives on the wild, wonderful, and sometimes weird world of video game news.

The Top 10 Most-Played Games On Steam Deck: August 2023 Edition

The Scott Pilgrim Anime Backlash, Explained

Image: Netflix

Scott Pilgrim Takes Off, the new animated series based on Bryan Lee O’Malley’s graphic novels, is out on Netflix. The eight-episode series reunites the voice cast of the 2010 live-action movie Scott Pilgrim vs. the World and is a hilarious blend of the series’ quick wit and well-measured pop culture references. All of this sounds like a recipe for success, right? Well, it’s a little more complicated. Read More

The New Division Game Has A Feature Every Game Should Steal

Image: Ubisoft / Kotaku

Ubisoft’s new The Division game isn’t even out yet, as it’s still in beta testing and won’t launch officially until 2024. But after trying the beta, I already want one feature from the upcoming game to become standard in every video game I play in the future. Read More

The Future Of ChatGPT Just Became A Circus [Update]

Photo: Justin Sullivan / Applle / Kotaku (Getty Images)

OpenAI is the research organization behind ChatGPT, the AI-generated chatbot that took the internet by storm last year for its capacity to have really weird conversations with tech journalists. It’s at the center of Microsoft’s big bet on generative AI tools transforming the world, gaming, and more, and it’s now at risk of imploding after its CEO, Sam Altman, was mysteriously ousted by the OpenAI board of directors and Twitch co-founder Emmett Shear was desperately recruited to replace him. Here’s all you really need to know about OpenAI to appreciate what a clusterfuck the last few days have been. Read More

Kotaku Asks: How Soon Is Too Soon For A Video Game Remaster Or Remake?

Screenshot: PlayStation / Naughty Dog

How much time has to pass before it becomes acceptable to remaster or even remake a game? 10 years? 15 years? What about three-ish years? Is that enough time between the original and the remaster? Well, that’s what’s happening early next year as Naughty Dog is remastering 2020’s The Last of Us Part II. Read More

I’m So Tired Of Crossover ‘Skins’ Cluttering Up Video Games

Image: Xbox / Epic Games / Bungie / Overkill Software / Kotaku

Another day, another big video game crossover. This time it’s Bungie’s online looter shooter, Destiny 2, adding Witcher 3-inspired armor to its digital store. Are you excited? I’m not. In reality, I’m just really tired of every brand mixing together, regardless of whether it makes sense or is needed, as if concocting the world’s worst stew. Read More

Admit It, You Don’t Understand Skill-Based Matchmaking (And Neither Do I)

Image: Kotaku / Asier Romero / Luis Molinero (Shutterstock)

Whenever a new blockbuster first-person shooter drops, gamers limber up so they can once again argue over how multiplayer matches get made and the algorithmic systems that determine who plays against whom and when. The recent release of Call of Duty: Modern Warfare III is no exception—not long after its multiplayer servers booted on November 10, players began flocking to Reddit, X (Twitter), and everywhere in between to complain about the quality (or perceived lack thereof) of Activision’s matchmaking. But, as with so many issues in the gaming industry, there’s a serious lack of nuance and true understanding at play here. Read More

I Can’t Miss The Last Of Us If It Won’t Leave

Image: Naughty Dog

Remember when it took us seven years to get a new The Last of Us game? Remember when there was even a question about whether or not we’d ever get a sequel to Naughty Dog’s post-apocalyptic action game because the ending was so intentionally ambiguous and thought-provoking?

Now, it seems we can’t go a year without being reminded that Sony thinks as many people should experience this series as possible, while folks associated with the HBO adaptation praise the game in ways that border on the absurd. Now, we’re getting a remaster of The Last of Us Part II, and it feels like we’re reaching peak Last of Us fatigue. Read More

This Modern Warfare 3 Gameplay Feature Spices Up A Weak Campaign

This Modern Warfare 3 Gameplay Feature Spices Up A Weak Campaign

Open Combat Missions are a fresh idea worth carrying over to future Call of Duty games.

[ad_2]

Kotaku Staff

Source link

November 25, 2023
WTF Fun Fact 13646 – Debating AI

[ad_1]

Debating AI might seem like a pointless venture – but you have a good chance of being told you’re right, even when you’re not.

Artificial intelligence, specifically large language models like ChatGPT, has shown remarkable capabilities in tackling complex questions. However, a study by The Ohio State University reveals an intriguing vulnerability: ChatGPT can be easily convinced that its correct answers are wrong. This discovery sheds light on the AI’s reasoning mechanisms and highlights potential limitations.

ChatGPT’s Inability to Uphold the Truth

Researchers conducted an array of debate-like conversations with ChatGPT, challenging the AI on its correct answers. The results were startling. Despite providing correct solutions initially, ChatGPT often conceded to invalid arguments posed by users, sometimes even apologizing for its supposedly incorrect answers. This phenomenon raises critical questions about the AI’s understanding of truth and its reasoning process.

AI’s prowess in complex reasoning tasks is well-documented. Yet, this study exposes a potential flaw: the inability to defend correct beliefs against trivial challenges. Boshi Wang, the study’s lead author, notes this contradiction. Despite AI’s efficiency in identifying patterns and rules, it struggles with simple critiques, similar to someone who copies information without fully comprehending it.

The Implications of Debating AI (and Winning)

The study’s findings imply significant concerns. For example, an AI system’s failure to uphold correct information in the face of opposition could lead to misinformation or wrong decisions, especially in critical fields like healthcare and criminal justice. The researchers aim to assess the safety of AI systems for human interaction, given their growing integration into various sectors.

Determining why ChatGPT fails to defend its correct answers is challenging due to the “black-box” nature of LLMs. The study suggests two possible causes: the base model’s lack of reasoning and truth understanding, and the influence of human feedback, which may teach the AI to yield to human opinion rather than stick to factual correctness.

Despite identifying this issue, solutions are not immediately apparent. Developing methods to enhance AI’s ability to maintain truth in the face of opposition will be crucial for its safe and effective application. The study marks an important step in understanding and improving the reliability of AI systems.

— WTF fun facts

Source: “ChatGPT often won’t defend its answers — even when it is right” — ScienceDaily

Related

[ad_2]

WTF

Source link

November 23, 2023
Unity Announces Big ‘AI’ Plans, Developers Have Concerns

[ad_1]

Video games engine provider Unity announced earlier today the introduction of two new machine-learning platforms, one of which in particular has developers and artists asking questions of the company that, at time of publishing, have yet to be answered.

The Week In Games: What’s Releasing Beyond Lara Croft

From Unity’s blog:

Today we’re announcing two new AI products: Unity Muse, an expansive platform for AI-driven assistance during creation, and Unity Sentis, which allows you to embed neural networks in your builds to enable previously unimaginable real-time experiences.

Muse is essentially just ChatGPT but for Unity specifically, and purports to let users ask questions about coding and resources and get instant answers. Sentis, however, is more concerning, as it “enables you to embed an AI model in the Unity Runtime for your game or application, enhancing gameplay and other functionality directly on end-user platforms.”

Because “AI” is a technology that in many cases is utterly reliant on work stolen from artists without consent or compensation, Unity’s announcement led to a lot of questions about Sentis, with particular focus on the tech’s ability to create stuff like images, models and animation. Scroll down past the announcement tweet, for example, and you’ll see a ton of variations of the same query:

just to jump on the train, which dataset y’all pull the art from???

Unity needs to be fully transparent about what ML models will be implemented, including the data they have been trained on. I don’t see any possible way ML, in current iterations, can be effective without training on countless ill gotten data.

REALLY concerning image generator stuff. What datasets?

Hi, what dataset was this trained on? Is this using artwork from artists without their permission? Animations? Materials? How was this AI trained?

You do realize that AI-created assets can’t be used commercially, so what was the rationale for adding this feature?

Which datasets were used in development of this? Did you negotiate & acquire all relevant licenses directly from copyright holders?

It’s a very specific question, one that at time of publishing Unity has yet to answer, either on Twitter or on the company’s forums (I’ve emailed the company asking the question specifically, and will update if I hear back). Those familiar with “AI”’s legal and copyright struggles can find the outline of an answer in this post by Unity employee TreyK-47, though, when he says you can’t use the tech as it exists today “for a current commercial or external project”.

Note that while there are clear dangers to jobs and the quality of games inherent in this push, those dangers are for the future; for the now, this looks (and sounds) like dogshit.

Experience the art of the possible | Unity AI

[ad_2]

Luke Plunkett

Source link

June 27, 2023