AI is developing all the time. Here are some picks from several articles what is expected to happen in AI and around it in 2025. Here are picks from various articles, the texts are picks from the article edited and in some cases translated for clarity.
AI in 2025: Five Defining Themes
https://news.sap.com/2025/01/ai-in-2025-defining-themes/
Artificial intelligence (AI) is accelerating at an astonishing pace, quickly moving from emerging technologies to impacting how businesses run. From building AI agents to interacting with technology in ways that feel more like a natural conversation, AI technologies are poised to transform how we work.
But what exactly lies ahead?
1. Agentic AI: Goodbye Agent Washing, Welcome Multi-Agent Systems
AI agents are currently in their infancy. While many software vendors are releasing and labeling the first “AI agents” based on simple conversational document search, advanced AI agents that will be able to plan, reason, use tools, collaborate with humans and other agents, and iteratively reflect on progress until they achieve their objective are on the horizon. The year 2025 will see them rapidly evolve and act more autonomously. More specifically, 2025 will see AI agents deployed more readily “under the hood,” driving complex agentic workflows.
In short, AI will handle mundane, high-volume tasks while the value of human judgement, creativity, and quality outcomes will increase.
2. Models: No Context, No Value
Large language models (LLMs) will continue to become a commodity for vanilla generative AI tasks, a trend that has already started. LLMs are drawing on an increasingly tapped pool of public data scraped from the internet. This will only worsen, and companies must learn to adapt their models to unique, content-rich data sources.
We will also see a greater variety of foundation models that fulfill different purposes. Take, for example, physics-informed neural networks (PINNs), which generate outcomes based on predictions grounded in physical reality or robotics. PINNs are set to gain more importance in the job market because they will enable autonomous robots to navigate and execute tasks in the real world.
Models will increasingly become more multimodal, meaning an AI system can process information from various input types.
3. Adoption: From Buzz to Business
While 2024 was all about introducing AI use cases and their value for organizations and individuals alike, 2025 will see the industry’s unprecedented adoption of AI specifically for businesses. More people will understand when and how to use AI, and the technology will mature to the point where it can deal with critical business issues such as managing multi-national complexities. Many companies will also gain practical experience working for the first time through issues like AI-specific legal and data privacy terms (compared to when companies started moving to the cloud 10 years ago), building the foundation for applying the technology to business processes.
4. User Experience: AI Is Becoming the New UI
AI’s next frontier is seamlessly unifying people, data, and processes to amplify business outcomes. In 2025, we will see increased adoption of AI across the workforce as people discover the benefits of humans plus AI.
This means disrupting the classical user experience from system-led interactions to intent-based, people-led conversations with AI acting in the background. AI copilots will become the new UI for engaging with a system, making software more accessible and easier for people. AI won’t be limited to one app; it might even replace them one day. With AI, frontend, backend, browser, and apps are blurring. This is like giving your AI “arms, legs, and eyes.”
5. Regulation: Innovate, Then Regulate
It’s fair to say that governments worldwide are struggling to keep pace with the rapid advancements in AI technology and to develop meaningful regulatory frameworks that set appropriate guardrails for AI without compromising innovation.
12 AI predictions for 2025
This year we’ve seen AI move from pilots into production use cases. In 2025, they’ll expand into fully-scaled, enterprise-wide deployments.
https://www.cio.com/article/3630070/12-ai-predictions-for-2025.html
This year we’ve seen AI move from pilots into production use cases. In 2025, they’ll expand into fully-scaled, enterprise-wide deployments.
1. Small language models and edge computing
Most of the attention this year and last has been on the big language models — specifically on ChatGPT in its various permutations, as well as competitors like Anthropic’s Claude and Meta’s Llama models. But for many business use cases, LLMs are overkill and are too expensive, and too slow, for practical use.
“Looking ahead to 2025, I expect small language models, specifically custom models, to become a more common solution for many businesses,”
2. AI will approach human reasoning ability
In mid-September, OpenAI released a new series of models that thinks through problems much like a person would, it claims. The company says it can achieve PhD-level performance in challenging benchmark tests in physics, chemistry, and biology. For example, the previous best model, GPT-4o, could only solve 13% of the problems on the International Mathematics Olympiad, while the new reasoning model solved 83%.
If AI can reason better, then it will make it possible for AI agents to understand our intent, translate that into a series of steps, and do things on our behalf, says Gartner analyst Arun Chandrasekaran. “Reasoning also helps us use AI as more of a decision support system,”
3. Massive growth in proven use cases
This year, we’ve seen some use cases proven to have ROI, says Monteiro. In 2025, those use cases will see massive adoption, especially if the AI technology is integrated into the software platforms that companies are already using, making it very simple to adopt.
“The fields of customer service, marketing, and customer development are going to see massive adoption,”
4. The evolution of agile development
The agile manifesto was released in 2001 and, since then, the development philosophy has steadily gained over the previous waterfall style of software development.
“For the last 15 years or so, it’s been the de-facto standard for how modern software development works,”
5. Increased regulation
At the end of September, California governor Gavin Newsom signed a law requiring gen AI developers to disclose the data they used to train their systems, which applies to developers who make gen AI systems publicly available to Californians. Developers must comply by the start of 2026.
There are also regulations about the use of deep fakes, facial recognition, and more. The most comprehensive law, the EU’s AI Act, which went into effect last summer, is also something that companies will have to comply with starting in mid-2026, so, again, 2025 is the year when they will need to get ready.
6. AI will become accessible and ubiquitous
With gen AI, people are still at the stage of trying to figure out what gen AI is, how it works, and how to use it.
“There’s going to be a lot less of that,” he says. But gen AI will become ubiquitous and seamlessly woven into workflows, the way the internet is today.
7. Agents will begin replacing services
Software has evolved from big, monolithic systems running on mainframes, to desktop apps, to distributed, service-based architectures, web applications, and mobile apps. Now, it will evolve again, says Malhotra. “Agents are the next phase,” he says. Agents can be more loosely coupled than services, making these architectures more flexible, resilient and smart. And that will bring with it a completely new stack of tools and development processes.
8. The rise of agentic assistants
In addition to agents replacing software components, we’ll also see the rise of agentic assistants, adds Malhotra. Take for example that task of keeping up with regulations.
Today, consultants get continuing education to stay abreast of new laws, or reach out to colleagues who are already experts in them. It takes time for the new knowledge to disseminate and be fully absorbed by employees.
“But an AI agent can be instantly updated to ensure that all our work is compliant with the new laws,” says Malhotra. “This isn’t science fiction.”
9. Multi-agent systems
Sure, AI agents are interesting. But things are going to get really interesting when agents start talking to each other, says Babak Hodjat, CTO of AI at Cognizant. It won’t happen overnight, of course, and companies will need to be careful that these agentic systems don’t go off the rails.
Companies such as Sailes and Salesforce are already developing multi-agent workflows.
10. Multi-modal AI
Humans and the companies we build are multi-modal. We read and write text, we speak and listen, we see and we draw. And we do all these things through time, so we understand that some things come before other things. Today’s AI models are, for the most part, fragmentary. One can create images, another can only handle text, and some recent ones can understand or produce video.
11. Multi-model routing
Not to be confused with multi-modal AI, multi-modal routing is when companies use more than one LLM to power their gen AI applications. Different AI models are better at different things, and some are cheaper than others, or have lower latency. And then there’s the matter of having all your eggs in one basket.
“A number of CIOs I’ve spoken with recently are thinking about the old ERP days of vendor lock,” says Brett Barton, global AI practice leader at Unisys. “And it’s top of mind for many as they look at their application portfolio, specifically as it relates to cloud and AI capabilities.”
Diversifying away from using just a single model for all use cases means a company is less dependent on any one provider and can be more flexible as circumstances change.
12. Mass customization of enterprise software
Today, only the largest companies, with the deepest pockets, get to have custom software developed specifically for them. It’s just not economically feasible to build large systems for small use cases.
“Right now, people are all using the same version of Teams or Slack or what have you,” says Ernst & Young’s Malhotra. “Microsoft can’t make a custom version just for me.” But once AI begins to accelerate the speed of software development while reducing costs, it starts to become much more feasible.
9 IT resolutions for 2025
https://www.cio.com/article/3629833/9-it-resolutions-for-2025.html
1. Innovate
“We’re embracing innovation,”
2. Double down on harnessing the power of AI
Not surprisingly, getting more out of AI is top of mind for many CIOs.
“I am excited about the potential of generative AI, particularly in the security space,”
3. And ensure effective and secure AI rollouts
“AI is everywhere, and while its benefits are extensive, implementing it effectively across a corporation presents challenges. Balancing the rollout with proper training, adoption, and careful measurement of costs and benefits is essential, particularly while securing company assets in tandem,”
4. Focus on responsible AI
The possibilities of AI grow by the day — but so do the risks.
“My resolution is to mature in our execution of responsible AI,”
“AI is the new gold and in order to truly maximize it’s potential, we must first have the proper guardrails in place. Taking a human-first approach to AI will help ensure our state can maintain ethics while taking advantage of the new AI innovations.”
5. Deliver value from generative AI
As organizations move from experimenting and testing generative AI use cases, they’re looking for gen AI to deliver real business value.
“As we go into 2025, we’ll continue to see the evolution of gen AI. But it’s no longer about just standing it up. It’s more about optimizing and maximizing the value we’re getting out of gen AI,”
6. Empower global talent
Although harnessing AI is a top objective for Morgan Stanley’s Wetmur, she says she’s equally committed to harnessing the power of people.
7. Create a wholistic learning culture
Wetmur has another talent-related objective: to create a learning culture — not just in her own department but across all divisions.
8. Deliver better digital experiences
Deltek’s Cilsick has her sights set on improving her company’s digital employee experience, believing that a better DEX will yield benefits in multiple ways.
Cilsick says she first wants to bring in new technologies and automation to “make things as easy as possible,” mirroring the digital experiences most workers have when using consumer technologies.
“It’s really about leveraging tech to make sure [employees] are more efficient and productive,”
“In 2025 my primary focus as CIO will be on transforming operational efficiency, maximizing business productivity, and enhancing employee experiences,”
9. Position the company for long-term success
Lieberman wants to look beyond 2025, saying another resolution for the year is “to develop a longer-term view of our technology roadmap so that we can strategically decide where to invest our resources.”
“My resolutions for 2025 reflect the evolving needs of our organization, the opportunities presented by AI and emerging technologies, and the necessity to balance innovation with operational efficiency,”
Lieberman aims to develop AI capabilities to automate routine tasks.
“Bots will handle common inquiries ranging from sales account summaries to HR benefits, reducing response times and freeing up resources for strategic initiatives,”
Not just hype — here are real-world use cases for AI agents
https://venturebeat.com/ai/not-just-hype-here-are-real-world-use-cases-for-ai-agents/
Just seven or eight months ago, when a customer called in to or emailed Baca Systems with a service question, a human agent handling the query would begin searching for similar cases in the system and analyzing technical documents.
This process would take roughly five to seven minutes; then the agent could offer the “first meaningful response” and finally begin troubleshooting.
But now, with AI agents powered by Salesforce, that time has been shortened to as few as five to 10 seconds.
Now, instead of having to sift through databases for previous customer calls and similar cases, human reps can ask the AI agent to find the relevant information. The AI runs in the background and allows humans to respond right away, Russo noted.
AI can serve as a sales development representative (SDR) to send out general inquires and emails, have a back-and-forth dialogue, then pass the prospect to a member of the sales team, Russo explained.
But once the company implements Salesforce’s Agentforce, a customer needing to modify an order will be able to communicate their needs with AI in natural language, and the AI agent will automatically make adjustments. When more complex issues come up — such as a reconfiguration of an order or an all-out venue change — the AI agent will quickly push the matter up to a human rep.
Open Source in 2025: Strap In, Disruption Straight Ahead
Look for new tensions to arise in the New Year over licensing, the open source AI definition, security and compliance, and how to pay volunteer maintainers.
https://thenewstack.io/open-source-in-2025-strap-in-disruption-straight-ahead/
The trend of widely used open source software moving to more restrictive licensing isn’t new.
In addition to the demands of late-stage capitalism and impatient investors in companies built on open source tools, other outside factors are pressuring the open source world. There’s the promise/threat of generative AI, for instance. Or the shifting geopolitical landscape, which brings new security concerns and governance regulations.
What’s ahead for open source in 2025?
More Consolidation, More Licensing Changes
The Open Source AI Debate: Just Getting Started
Security and Compliance Concerns Will Rise
Paying Maintainers: More Cash, Creativity Needed
Kyberturvallisuuden ja tekoälyn tärkeimmät trendit 2025
https://www.uusiteknologia.fi/2024/11/20/kyberturvallisuuden-ja-tekoalyn-tarkeimmat-trendit-2025/
1. Cyber infrastructure will be centered on a single, unified security platform
2. Big data will give an edge against new entrants
3. AI’s integrated role in 2025 means building trust, governance engagement, and a new kind of leadership
4. Businesses will adopt secure enterprise browsers more widely
5. AI’s energy implications will be more widely recognized in 2025
6. Quantum realities will become clearer in 2025
7. Security and marketing leaders will work more closely together
Presentation: For 2025, ‘AI eats the world’.
https://www.ben-evans.com/presentations
Just like other technologies that have gone before, such as cloud and cybersecurity automation, right now AI lacks maturity.
https://www.securityweek.com/ai-implementing-the-right-technology-for-the-right-use-case/
If 2023 and 2024 were the years of exploration, hype and excitement around AI, 2025 (and 2026) will be the year(s) that organizations start to focus on specific use cases for the most productive implementations of AI and, more importantly, to understand how to implement guardrails and governance so that it is viewed as less of a risk by security teams and more of a benefit to the organization.
Businesses are developing applications that add Large Language Model (LLM) capabilities to provide superior functionality and advanced personalization
Employees are using third party GenAI tools for research and productivity purposes
Developers are leveraging AI-powered code assistants to code faster and meet challenging production deadlines
Companies are building their own LLMs for internal use cases and commercial purposes.
AI is still maturing
However, just like other technologies that have gone before, such as cloud and cybersecurity automation, right now AI lacks maturity. Right now, we very much see AI in this “peak of inflated expectations” phase and predict that it will dip into the “trough of disillusionment”, where organizations realize that it is not the silver bullet they thought it would be. In fact, there are already signs of cynicism as decision-makers are bombarded with marketing messages from vendors and struggle to discern what is a genuine use case and what is not relevant for their organization.
There is also regulation that will come into force, such as the EU AI Act, which is a comprehensive legal framework that sets out rules for the development and use of AI.
AI certainly won’t solve every problem, and it should be used like automation, as part of a collaborative mix of people, process and technology. You simply can’t replace human intuition with AI, and many new AI regulations stipulate that human oversight is maintained.
7 Splunk Predictions for 2025
https://www.splunk.com/en_us/form/future-predictions.html
AI: Projects must prove their worth to anxious boards or risk defunding, and LLMs will go small to reduce operating costs and environmental impact.
OpenAI, Google and Anthropic Are Struggling to Build More Advanced AI
Three of the leading artificial intelligence companies are seeing diminishing returns from their costly efforts to develop newer models.
https://www.bloomberg.com/news/articles/2024-11-13/openai-google-and-anthropic-are-struggling-to-build-more-advanced-ai
Sources: OpenAI, Google, and Anthropic are all seeing diminishing returns from costly efforts to build new AI models; a new Gemini model misses internal targets
It Costs So Much to Run ChatGPT That OpenAI Is Losing Money on $200 ChatGPT Pro Subscriptions
https://futurism.com/the-byte/openai-chatgpt-pro-subscription-losing-money?fbclid=IwY2xjawH8epVleHRuA2FlbQIxMQABHeggEpKe8ZQfjtPRC0f2pOI7A3z9LFtFon8lVG2VAbj178dkxSQbX_2CJQ_aem_N_ll3ETcuQ4OTRrShHqNGg
In a post on X-formerly-Twitter, CEO Sam Altman admitted an “insane” fact: that the company is “currently losing money” on ChatGPT Pro subscriptions, which run $200 per month and give users access to its suite of products including its o1 “reasoning” model.
“People use it much more than we expected,” the cofounder wrote, later adding in response to another user that he “personally chose the price and thought we would make some money.”
Though Altman didn’t explicitly say why OpenAI is losing money on these premium subscriptions, the issue almost certainly comes down to the enormous expense of running AI infrastructure: the massive and increasing amounts of electricity needed to power the facilities that power AI, not to mention the cost of building and maintaining those data centers. Nowadays, a single query on the company’s most advanced models can cost a staggering $1,000.
Tekoäly edellyttää yhä nopeampia verkkoja
https://etn.fi/index.php/opinion/16974-tekoaely-edellyttaeae-yhae-nopeampia-verkkoja
A resilient digital infrastructure is critical to effectively harnessing telecommunications networks for AI innovations and cloud-based services. The increasing demand for data-rich applications related to AI requires a telecommunications network that can handle large amounts of data with low latency, writes Carl Hansson, Partner Solutions Manager at Orange Business.
AI’s Slowdown Is Everyone Else’s Opportunity
Businesses will benefit from some much-needed breathing space to figure out how to deliver that all-important return on investment.
https://www.bloomberg.com/opinion/articles/2024-11-20/ai-slowdown-is-everyone-else-s-opportunity
Näin sirumarkkinoilla käy ensi vuonna
https://etn.fi/index.php/13-news/16984-naein-sirumarkkinoilla-kaey-ensi-vuonna
The growing demand for high-performance computing (HPC) for artificial intelligence and HPC computing continues to be strong, with the market set to grow by more than 15 percent in 2025, IDC estimates in its recent Worldwide Semiconductor Technology Supply Chain Intelligence report.
IDC predicts eight significant trends for the chip market by 2025.
1. AI growth accelerates
2. Asia-Pacific IC Design Heats Up
3. TSMC’s leadership position is strengthening
4. The expansion of advanced processes is accelerating.
5. Mature process market recovers
6. 2nm Technology Breakthrough
7. Restructuring the Packaging and Testing Market
8. Advanced packaging technologies on the rise
2024: The year when MCUs became AI-enabled
https://www-edn-com.translate.goog/2024-the-year-when-mcus-became-ai-enabled/?fbclid=IwZXh0bgNhZW0CMTEAAR1_fEakArfPtgGZfjd-NiPd_MLBiuHyp9qfiszczOENPGPg38wzl9KOLrQ_aem_rLmf2vF2kjDIFGWzRVZWKw&_x_tr_sl=en&_x_tr_tl=fi&_x_tr_hl=fi&_x_tr_pto=wapp
The AI party in the MCU space started in 2024, and in 2025, it is very likely that there will be more advancements in MCUs using lightweight AI models.
Adoption of AI acceleration features is a big step in the development of microcontrollers. The inclusion of AI features in microcontrollers started in 2024, and it is very likely that in 2025, their features and tools will develop further.
Just like other technologies that have gone before, such as cloud and cybersecurity automation, right now AI lacks maturity.
https://www.securityweek.com/ai-implementing-the-right-technology-for-the-right-use-case/
If 2023 and 2024 were the years of exploration, hype and excitement around AI, 2025 (and 2026) will be the year(s) that organizations start to focus on specific use cases for the most productive implementations of AI and, more importantly, to understand how to implement guardrails and governance so that it is viewed as less of a risk by security teams and more of a benefit to the organization.
Businesses are developing applications that add Large Language Model (LLM) capabilities to provide superior functionality and advanced personalization
Employees are using third party GenAI tools for research and productivity purposes
Developers are leveraging AI-powered code assistants to code faster and meet challenging production deadlines
Companies are building their own LLMs for internal use cases and commercial purposes.
AI is still maturing
AI Regulation Gets Serious in 2025 – Is Your Organization Ready?
While the challenges are significant, organizations have an opportunity to build scalable AI governance frameworks that ensure compliance while enabling responsible AI innovation.
https://www.securityweek.com/ai-regulation-gets-serious-in-2025-is-your-organization-ready/
Similar to the GDPR, the EU AI Act will take a phased approach to implementation. The first milestone arrives on February 2, 2025, when organizations operating in the EU must ensure that employees involved in AI use, deployment, or oversight possess adequate AI literacy. Thereafter from August 1 any new AI models based on GPAI standards must be fully compliant with the act. Also similar to GDPR is the threat of huge fines for non-compliance – EUR 35 million or 7 percent of worldwide annual turnover, whichever is higher.
While this requirement may appear manageable on the surface, many organizations are still in the early stages of defining and formalizing their AI usage policies.
Later phases of the EU AI Act, expected in late 2025 and into 2026, will introduce stricter requirements around prohibited and high-risk AI applications. For organizations, this will surface a significant governance challenge: maintaining visibility and control over AI assets.
Tracking the usage of standalone generative AI tools, such as ChatGPT or Claude, is relatively straightforward. However, the challenge intensifies when dealing with SaaS platforms that integrate AI functionalities on the backend. Analysts, including Gartner, refer to this as “embedded AI,” and its proliferation makes maintaining accurate AI asset inventories increasingly complex.
Where frameworks like the EU AI Act grow more complex is their focus on ‘high-risk’ use cases. Compliance will require organizations to move beyond merely identifying AI tools in use; they must also assess how these tools are used, what data is being shared, and what tasks the AI is performing. For instance, an employee using a generative AI tool to summarize sensitive internal documents introduces very different risks than someone using the same tool to draft marketing content.
For security and compliance leaders, the EU AI Act represents just one piece of a broader AI governance puzzle that will dominate 2025.
The next 12-18 months will require sustained focus and collaboration across security, compliance, and technology teams to stay ahead of these developments.
The Global Partnership on Artificial Intelligence (GPAI) is a multi-stakeholder initiative which aims to bridge the gap between theory and practice on AI by supporting cutting-edge research and applied activities on AI-related priorities.
https://gpai.ai/about/#:~:text=The%20Global%20Partnership%20on%20Artificial,activities%20on%20AI%2Drelated%20priorities.
289 Comments
Tomi Engdahl says:
DeepSeek R1’s bold bet on reinforcement learning: How it outpaced OpenAI at 3% of the cost
https://venturebeat.com/ai/deepseek-r1s-bold-bet-on-reinforcement-learning-how-it-outpaced-openai-at-3-of-the-cost/
Tomi Engdahl says:
https://wonderfulengineering.com/ai-cant-even-spell-strawberry-here-is-why/
Tomi Engdahl says:
DeepSeek-R1: The Open-Source AI Changing the Game in Technology
https://medium.com/@soaltinuc/deepseek-r1-the-open-source-ai-changing-the-game-in-technology-15132b99b9d7
One of the most important features of DeepSeek-R1 is that it is open source. It is released under the MIT license, making its technology and research publicly available. They also released the research paper, and people are replicating the research and trying the methodology in new places.
Paper: https://arxiv.org/pdf/2501.12948
Both for general replication: Open R1 is replicating the whole training process. https://github.com/huggingface/open-r1
And TinyZero is replicating the method for specific purposes: https://github.com/Jiayi-Pan/TinyZero, which is optimized for countdown and multiplication tasks and was trained for just $30.
Also, DeepSeek-R1 is very cost-effective: training the model is 95% less expensive than its competitors, and its API costs just $0.55 per million tokens — 2% of the cost of OpenAI O1. This new cost is on a disruptive level, making the use of reasoning models feasible for most businesses.
What’s more is they even have distilled models with similar reasoning capabilities that run on much smaller hardware. Basically, an O1-mini on your computer. They released a wide range of distilled models ranging from 1.5B to 70B that could provide GPT-4 level performance for most tasks on customer GPUs.
The technical architecture of DeepSeek-R1 is quite impressive. It uses a Mixture-of-Experts (MoE) setup; though it has a total of 671 billion parameters, only 37 billion are activated per task. This allows the model to operate efficiently without compromising performance, as it can draw upon specialized parameters for particular tasks and capabilities.
The training is also highly novel. While most LLMs are based mainly on SFT, DeepSeek-R1 relies more on RL and a minimum of SFT.
This includes:
DeepSeek-R1-Zero: For the first time, the model is trained with RL only, allowing the model to learn complicated reasoning skills autonomously, such as self-verification and Chain-of-Thought (CoT) reasoning.
DeepSeek-R1: Improved from R1-Zero by using a curated “cold start” dataset and multi-stage RL to fine-tune readability, mitigating issues such as language mixing.
GRPO: A new algorithm enabling the model to enhance its outputs by generating multiple responses, evaluating them for accuracy and format, and reinforcing the best results.
What they have done for R1 is basically:
Started with custom, carefully curated data for the cold start phase, which allowed the model to perform better in the RL stage and have specialized tokens for thinking.
Ran GRPO RL algorithm on language consistency and accuracy rewards to enhance reasoning. This is the secret ingredient that provides reasoning capabilities.
Performed rejection sampling and supervised fine-tuning with data filtered from R1 checkpoint model outputs and samples from Deepseek v3 to provide data from other domains like writing and general knowledge. This helps the model to be useful and knowledgeable on top of its reasoning capabilities.
Ran RL algorithm for different scenarios with reward models generated for data. This improves helpfulness and harmlessness.
Tomi Engdahl says:
DeepSeek-R1’s strengths lie in problem-solving and reasoning, while OpenAI’s O1–1217 slightly leads in general knowledge and coding tasks.
Tomi Engdahl says:
Key Takeaways
DeepSeek-R1 is an open-source model that rivals top proprietary AI systems.
It’s significantly cheaper to train and use, lowering barriers to entry.
Some of the innovative training methods include reinforcement learning and MoE architecture.
The model performs fantastically on reasoning-intensive tasks such as mathematics and software engineering.
It seems like open source can keep up the pace, and the gap has closed to three months.
Tomi Engdahl says:
https://simonwillison.net/2025/Jan/25/openai-canvas-gets-a-huge-upgrade/
Tomi Engdahl says:
https://www.techspot.com/news/106503-microsoft-365-copilot-rollout-sparks-backlash-over-price.html
Tomi Engdahl says:
https://github.com/deepseek-ai/DeepSeek-V3
Tomi Engdahl says:
https://www.boredpanda.com/zero-points-kids-using-ai-teacher-proof/
Tomi Engdahl says:
Vector Databases: The Foundation of AI Agent Innovation
Organizations must dedicate time and resources specifically to AI and vector database technology.
https://thenewstack.io/vector-databases-the-foundation-of-ai-agent-innovation/
Vector databases will be key to building AI agents, the “next frontier of generative AI,” according to McKinsey. “We are beginning an evolution from knowledge-based, gen-AI-powered tools — say, chatbots that answer questions and generate content — to gen AI-enabled ‘agents’ that use foundation models to execute complex, multistep workflows across a digital world. In short, the technology is moving from thought to action,” a recent McKinsey Quarterly article states.
Deloitte predicts that 25% of enterprises using GenAI will deploy AI agents in 2025, growing to 50% by 2027. “While early adopters will grapple with complexities and challenges, the vision is compelling enough for organizations to take proactive steps to prepare themselves now for adoption,” according to Deloitte Global’s 2025 Predictions Report, Generative AI: Paving the Way for a Transformative Future in Technology, Media, and Telecommunications. “This evolution will enable AI agents to tackle a broader range of applications, providing businesses with valuable tools to drive the productivity of knowledge workers and efficiency gains in workflows of all kinds.”
Tomi Engdahl says:
https://www.mckinsey.com/capabilities/mckinsey-digital/our-insights/why-agents-are-the-next-frontier-of-generative-ai
Tomi Engdahl says:
https://www.marktechpost.com/2025/01/24/llasa-3b-a-llama-3-2b-fine-tuned-text-to-speech-model-with-ultra-realistic-audio-emotional-expressiveness-and-multilingual-support/
Tomi Engdahl says:
Tiny DeepSeek 1.5B Models Run on $249 NVIDIA Jetson Nano
https://www.nextbigfuture.com/2025/01/tiny-deepseek-1-5b-models-run-on-99-nvidia-jetson-nano.html
Youtuber, Ominous Industries, ran a couple of versions of the DeepSeek R1 1.5B of models running locally on the NVIDIA Jetson Nan. The newly released distilled DeepSeek models were explroed. The DeepSeek R1 1.5B model delivers impressive performance with plenty of room to spare on the Jetson.
He shows the installation process, followed by a series of tests, including a Python reasoning test. He compared the DeepSeek R1 distilled 1.5B models to the Mini Llama 3.1 1B model. He set up and tested the FP16 version of the DeepSeek R1 Distilled 1.5B Qwen model, running seamlessly in WebUI with an Ollama backend.
Tomi Engdahl says:
https://www.marktechpost.com/2025/01/25/meta-ai-releases-the-first-stable-version-of-llama-stack-a-unified-platform-transforming-generative-ai-development-with-backward-compatibility-safety-and-seamless-multi-environment-deployment/
Tomi Engdahl says:
https://www.tomshardware.com/tech-industry/artificial-intelligence/meta-to-build-2gw-data-center-with-over-1-3-million-nvidia-ai-gpus-invest-usd65b-in-ai-in-2025
Tomi Engdahl says:
https://puheenvuoro.uusisuomi.fi/riikka-tanner/jos-tekoaly-tekee-meista-tehokkaampia-mutta-vahemman-kriittisia-ajattelijoita-olemmeko-todella-voittaneet/
Tomi Engdahl says:
https://hackaday.com/2025/01/26/prompt-injection-tricks-ai-into-downloading-and-executing-malware/
Tomi Engdahl says:
https://www.cnet.com/tech/services-and-software/deepseek-strikes-again-with-ai-image-generator-janus-pro/#google_vignette
Tomi Engdahl says:
https://hackernoon.com/a-basic-knowledge-of-python-can-help-you-build-your-own-machine-learning-model
Tomi Engdahl says:
https://www.pcgamer.com/software/ai/ais-sputnik-moment-china-based-deepseeks-open-source-models-may-be-a-real-threat-to-the-dominance-of-openai-meta-and-nvidia/
Tomi Engdahl says:
Fully open reproduction of DeepSeek-R1
https://github.com/huggingface/open-r1
Tomi Engdahl says:
Sam Altman breaks silence on new ChatGPT rival DeepSeek after Chinese start up wiped $1,000,000,000,000 off stocks
Altman has praised the Chinese ChatGPT rival
https://www.uniladtech.com/news/ai/sam-altman-breaks-silence-chinese-ai-startup-deepseek-795270-20250128
OpenAI CEO Sam Altman has offered a response to the chaos that Chinese AI startup DeepSeek has enabled, as the ChatGPT rival has caused the stock market to drop by over $1,000,000,000,000.
Much of the AI and technology world has been sent into disarray after the release of DeepSeek, a ChatGPT-like open source AI technology that was made by a Chinese startup on a fraction of the budget of its western counterparts.
Reports indicate that it was developed with a budget of just $6,000,000 using Nvidia’s H800 chips, and it has seen overall losses of around $1 trillion in the stock market with Nvidia especially suffering the biggest value drop in history.
US President Donald Trump has declared that this should be a ‘wakeup call’ for the American AI industry, arguing that they should now be “laser-focused on competing to win” in light of DeepSeek’s almost revolutionary developments.
Altman outlined on X: “DeepSeek’s R1 is an impressive model, particularly around what they’re able to deliver for the price. We will obviously deliver much better models and also it’s legit invigorating to have a new competitor! We will pull up some releases.”
It’s currently unclear what exactly these new releases and models will be or include, but it does arguably show that Altman isn’t necessarily as threatened as many might have expected, and welcomes the push in the industry. Competition breeds innovation, after all.
Altman continued on to articulate that OpenAI is “excited to continue to execute on our research roadmap and believe more compute is more important now than ever before to succeed at our mission,” offering almost an inverse of what DeepSeek’s low-budget approach strives for.
Finally, Altman reiterates his goal for achieving artificial general intelligence, outlining that he “look[s] forward to bringing you all AGI and beyond.”
“How about open sourcing your models,” questions one user, whereas another adds that they’re “looking forward to a monthly subscription price reduce” that isn’t confirmed to be on its way.
What remains clear is that Altman and many other US AI giants have the support of the government, as President Trump only recently forged a $500,000,000 AI partnership that would look to instantly create 100,000 American jobs and even create personalized cancer vaccinations in the future.
Tomi Engdahl says:
https://theindependent.sg/meta-sets-up-4-war-rooms-to-tackle-deepseeks-ai-breakthrough-as-zuckerberg-plans-s87-9b-investment/?fbclid=IwY2xjawIHUTtleHRuA2FlbQIxMQABHYUnA8ewpa3t_xeLZIuQbuyj6a7HeJb_FSIQR0fo5_PmCYS89BHhybNT5Q_aem_hTApJFOM6qfE7YvhzRrhkA
Tomi Engdahl says:
Would you allow to run commands directly suggested by LLM on your personal Linux or macOS desktop? Ready to give an LLM the keys to your computer?
Tomi Engdahl says:
Deepseek-tekoäly myönsi yllättäen Kiinan viranomaisten vakoilun
https://www.is.fi/digitoday/art-2000010997729.html
Tomi Engdahl says:
Google Please Make Gemini Good at Coding | Gemini 1.5 and 2.0 Review
https://www.youtube.com/watch?v=L3yuif2mh9s
Hey everyone! In this video, I dive deep into my experience testing Google Gemini for coding tasks over 10 hours. As a huge fan of AI tools, I was excited to see how Gemini could handle everything from basic HTML to SQL and Python. But… it didn’t quite live up to my expectations.
From internal server errors to incomplete responses and super long response times, I encountered a lot of hurdles. Don’t get me wrong—I’m still rooting for Gemini to improve, but right now, it’s not quite ready to be your go-to coding assistant.
In this video, I’ll walk you through:
My testing process and the tasks I gave Gemini
The issues I ran into (and how often they happened)
Why I still believe in Gemini’s potential
What needs to improve for it to compete with other AI coding tools
Tomi Engdahl says:
Financial Times:
SemiAnalysis: DeepSeek has spent “well over $500M on GPUs over the history of the company”; TechInsights says it doesn’t see DeepSeek as “a big hit to Nvidia” — Short sellers profit as US chipmaker loses nearly $600bn in market value on Monday
https://www.ft.com/content/ee83c24c-9099-42a4-85c9-165e7af35105
Bloomberg:
DeepSeek says it used Nvidia H800 chips, available in China until October 2023, to train R1, suggesting future models could be hampered by US export controls
DeepSeek’s AI Model Tests Limits of US Restrictions on Nvidia Chips
https://www.bloomberg.com/news/articles/2025-01-27/deepseek-s-ai-model-tests-limits-of-us-curbs-on-nvidia-chips
Nvidia calls R1 an ‘excellent’ advance that meets US limits
Trump says release should be a ‘wake-up call’ for US companies
Tomi Engdahl says:
OpenAI says it has evidence suggesting Chinese AI startup DeepSeek used its proprietary models to train a competing open-source system through “distillation,” a technique where smaller models learn from larger ones’ outputs.
The San Francisco-based company, along with partner Microsoft, blocked suspected DeepSeek accounts from accessing its API last year after detecting potential terms of service violations. DeepSeek’s R1 reasoning model has achieved comparable results to leading U.S. models despite claiming minimal resources.
Read more of this story at Slashdot.
OpenAI Says It Has Evidence DeepSeek Used Its Model To Train Competitor
https://m.slashdot.org/story/438145
Tomi Engdahl says:
OpenAI has evidence that its models helped train China’s DeepSeekOh, the irony.
https://www.theverge.com/news/601195/openai-evidence-deepseek-distillation-ai-data
Chinese artificial intelligence company DeepSeek disrupted Silicon Valley with the release of cheaply developed AI models that compete with flagship offerings from OpenAI — but the ChatGPT maker suspects they were built upon OpenAI data.
OpenAI and Microsoft are investigating whether the Chinese rival used OpenAI’s API to integrate OpenAI’s AI models into DeepSeek’s own models, according to Bloomberg. The outlet’s sources said Microsoft security researchers detected that large amounts of data were being exfiltrated through OpenAI developer accounts in late 2024, which the company believes are affiliated with DeepSeek.
OpenAI told the Financial Times that it found evidence linking DeepSeek to the use of distillation — a common technique developers use to train AI models by extracting data from larger, more capable ones. It’s an efficient way to train smaller models at a fraction of the more than $100 million that OpenAI spent to train GPT-4. While developers can use OpenAI’s API to integrate its AI with their own applications, distilling the outputs to build rival models is a violation of OpenAI’s terms of service. OpenAI has not provided details of the evidence it found.
The situation is rich with irony. After all, it was OpenAI that made huge leaps with its GPT model by sucking down the entirety of the written web without consent.
President Donald Trump’s artificial intelligence czar David Sacks said “it is possible” that IP theft had occurred. “There’s substantial evidence that what DeepSeek did here is they distilled knowledge out of OpenAI models and I don’t think OpenAI is very happy about this,” Sacks told Fox News on Tuesday.
“We know PRC (China) based companies — and others — are constantly trying to distill the models of leading US AI companies,”
Tomi Engdahl says:
You mean they used AI to train another AI to cut down time energy and processing to achieve a better cheeper product
Tomi Engdahl says:
Oh dang the plagerism machine got plagerized oh no
“Their plagiarism machine plagiarized our plagiarism machine!”
Tomi Engdahl says:
Mark Zuckerberg says Meta isn’t worried about DeepSeek‘I continue to think that investing very heavily in CapEx and infra is going to be a strategic advantage over time.’
https://www.theverge.com/news/602233/meta-deepseek-zuckerberg-earings-q4?fbclid=IwZXh0bgNhZW0CMTEAAR2bliR0xlMRm0OhvxEuP6xYZ5cvj3RIg_wHODRCZWBonGCPjdWif4MSgGM_aem_AiQ2-RDA9zUt6cXR7835-A
Nearly everyone seems to be suddenly freaking out about the rise of DeepSeek. Meta isn’t worried, though.
That was CEO Mark Zuckerberg’s message to investors during his company’s fourth-quarter earnings call on Wednesday. During the Q&A portion of the call with Wall Street analysts, Zuckerberg fielded multiple questions about DeepSeek’s impressive AI models and what the implications are for Meta’s AI strategy. He said that what DeepSeek was able to accomplish with relatively little money has “only strengthened our conviction that this is the right thing to be focused on.”
Tomi Engdahl says:
A closer look at DeepSeek’s AI affair and what it means for Nvidia
DeepSeek’s AI stunner and the future of Nvidia
https://www.edn.com/deepseeks-ai-stunner-and-the-future-of-nvidia/?fbclid=IwZXh0bgNhZW0CMTEAAR2slDQb8puA58TMyE91xu5Ug8NO4XFj9JoyZbKlnokv40lij9ukqBhIubk_aem_hxq7A3gST9shoMmiRcIqMQ#google_vignette
After the release of DeepSeek’s R1, a reasoning LLM that matches the performance of OpenAI’s latest o1 model, trade media is abuzz with speculations about the future of artificial intelligence (AI). Has the AI bubble burst? Is it the end of Nvidia’s spectacular AI ride?
As Nvidia Stock Drops, Has The AI Chip Bubble Finally Burst?
https://www.eetimes.com/as-nvidia-stock-drops-15-has-the-ai-chip-bubble-finally-burst/
Well, that was not quite the effect Silicon Valley, and President Donald Trump, were expecting when they announced the $500-billion Stargate AI infrastructure project last week.
In a move that, timing wise, seems unlikely to be coincidental, Chinese AI lab DeepSeek released DeepSeek-R1, a reasoning LLM that matches the performance of OpenAI’s latest o1 model, on Trump’s inauguration day, Jan. 20th. DeepSeek-R1 is a fine-tuned version of DeepSeek-V3, which was pre-trained using about $5.5 million of compute, or around 10-20× less than other comparably-sized LLMs.
The market’s reaction to this was to wipe $500 billion off the market cap of Nvidia (stock dropped 17% in less than a day).
DeepSeek techniques
DeepSeek, a spin-out from AI-driven Chinese hedge fund High-Flyer AI, trained V3 on an extremely modest cluster of 2,048 Nvidia H800 GPUs. The H800 is a cut-down version of the market-leading H100 for the Chinese market, which was designed to skirt the U.S. export regulations at the time. The H800s are compute capped and have reduced chip-to-chip communication bandwidth, which is vital for training LLMs.
For V3, a 671B MoE model that activates about 37B parameters on each forward pass, DeepSeek’s paper says they used 2.788 million GPU hours to pre-train on 14.8 trillion tokens. On their cluster of 2,048 GPUs, that would have taken 56 days, and at $2 per GPU-hour, the cost is estimated at $5.5 million. This figure is for the pre-training stage of V3 only.
For comparison, Llama 3.1-405B was trained in 30.8 million GPU hours, on 16,000 H100s, with slightly more data. The cost difference is still about a factor of 10.
How did DeepSeek train so efficiently? According to their paper, they have several tricks up their sleeves. One of the biggest seems to be DualPipe, a parallel-pipelining algorithm they invented, which successfully overlaps compute and communication in such a way that most of the communication overhead is hidden.
“As the model scales up […] we can still employ fine-grained experts across nodes while achieving a near-zero all-to-all communication overhead,” according to the paper. The company also developed all-to-all communication kernels (at the PTX level, a level below CUDA code) that better utilize Infiniband and NVLink bandwidth, and tinkered with the memory footprint to avoid having to use tensor parallelism at all. The net result is to hide communication bottlenecks. The company also dropped precision to FP8 in as many places as possible, among other techniques.
To get from V3 to R1, DeepSeek reportedly used additional reinforcement learning and supervised fine-tuning stages to improve the model’s reasoning capabilities.
Realistic cost
It should be noted that the $5.5 million figure is projected (not a real dollar spend) for one training run, one time. Development of the V3 model surely took months or years of expensive research and development, including failed training runs that are not counted here.
DeepSeek or its parent company had to invest in presumably many thousands of whatever GPUs they could get their hands on. Huge investment has been required to get to this stage, even if it is not on the same scale as OpenAI or U.S. infrastructure.
DeepSeek-R1 did not develop in a vacuum—it would not have been developed without the example of a cutting-edge reasoning model as a target, with OpenAI’s o1 the most well-known example. It could be argued that R1 is less innovation, more replication, based on frontiers already broken by the big U.S. AI labs at their expense. The techniques to get there may be innovative, but the result is still analogous to o1.
There is also reportedly a more literal link. The FT reports that OpenAI has evidence its o1 model was used to create synthetic training data for R1—in other words, OpenAI is alleging DeepSeek used its API to effectively copy its model. This is a well-known technique called distillation in which smaller LLMs are made more efficient by training them to copy outputs of bigger LLMs.
This is a great “cheat” but obviously when used by another company, it is akin to riding OpenAI’s coat-tails, or less charitably, stealing its expensively-developed IP. While OpenAI’s API is not restricted by U.S. export regulations currently, using the data in this way is expressly prohibited by OpenAI’s terms of service.
Nvidia markets
So, is the result that people will buy more Nvidia chips or less?
One could argue that Nvidia could suffer in the short term—if performance advantages resulting from this work mean companies hang onto their Hopper-generation GPUs for longer
Nvidia could also suffer the effects of newly-introduced U.S. export restrictions that substantially limit its overseas markets, unrelated to DeepSeek and these new developments.
In the longer term, though, it seems likely that as AI grows, it will need more chips.
Suddenly, X (Twitter) and LinkedIn are filled with armchair experts on Jevon’s Paradox: the phenomenon where as a technology gets cheaper or more efficient, more of that technology will be sold, not less. In this case, the argument goes: as AI gets cheaper to train, more AIs will be trained and the resulting growth and proliferation in AI technology will drive the market for AI chips.
Let u not forget that DeepSeek’s specific training techniques are applicable only to Nvidia GPU clusters, effectively proving you still need Nvidia chips to train this efficiently. No doubt the big U.S. AI labs are working hard to replicate and implement similar techniques as we speak. They will need Nvidia chips to do so.
While Nvidia faces challenges from competitors and the U.S. government, it still has its software moat and it still has a huge installed base. It is still selling more chips than it can make. Anyone working out ways to use GPUs more efficiently will ultimately help the propagation of AI, which will help Nvidia sell more chips.
Tomi Engdahl says:
Microsoft:
Microsoft reports Q2 revenue up 12% YoY to $69.6B, Microsoft 365 Commercial products and cloud services revenue up 15% YoY, and net income up 10% YoY to $24.1B — Microsoft Cloud and AI Strength Drives Second Quarter Results — REDMOND, Wash. — January 29, 2025 — Microsoft Corp. today announced …
https://www.microsoft.com/en-us/Investor/earnings/FY-2025-Q2/press-release-webcast
Jordan Novet / CNBC:
Microsoft reports Q2 Intelligent Cloud revenue up 19% YoY to $25.5B, vs. $25.8B est., with Azure and other cloud services revenue up 31%, down from 33% in Q1 — Microsoft shares tumbled as much as 5% in extended trading Wednesday after the software company issued fiscal second-quarter results …
Microsoft shares slip on weak quarterly revenue guidance
https://www.cnbc.com/2025/01/29/microsoft-msft-q2-earnings-report-2025.html
Todd Bishop / GeekWire:
Satya Nadella says Microsoft’s AI business has surpassed an annual revenue run rate of $13B, up 175% YoY, as the company faces new scrutiny over AI spending
Microsoft’s AI revenue run rate reaches $13B annually as tech giant tops earnings expectations
https://www.geekwire.com/2025/microsoft-earnings-2/
Tomi Engdahl says:
Meta:
Meta reports Q4 revenue up 21% YoY to $48.4B, net income up 49% YoY to $20.8B, and family daily active people up 5% YoY to 3.35B on average for December 2024 — Meta Platforms, Inc. (Nasdaq: META) today reported financial results for the quarter and full year ended December 31, 2024.
https://investor.atmeta.com/investor-news/press-release-details/2025/Meta-Reports-Fourth-Quarter-and-Full-Year-2024-Results/default.aspx
Charles Rollet / TechCrunch:
Mark Zuckerberg says spending heavily on AI infrastructure is a “strategic advantage” and vows Meta will invest “hundreds of billions” in AI over the long term — U.S. markets panicked on Monday over speculation that DeepSeek’s AI models would crush demand for GPUs, with Nvidia’s stock dropping almost 20%.
Zuck shrugs off DeepSeek, vows to spend hundreds of billions on AI
https://techcrunch.com/2025/01/29/zuck-shrugs-off-deepseek-vows-to-spend-hundreds-of-billions-on-ai/
Tomi Engdahl says:
Wired:
Wiz: DeepSeek left one of its critical databases exposed, leaking more than 1M records including system logs, user prompt submissions, and users’ API keys — China-based DeepSeek has exploded in popularity, drawing greater scrutiny. Case in point: Security researchers found more than 1 million records …
Exposed DeepSeek Database Revealed Chat Prompts and Internal Data
China-based DeepSeek has exploded in popularity, drawing greater scrutiny. Case in point: Security researchers found more than 1 million records, including user data and API keys, in an open database.
https://www.wired.com/story/exposed-deepseek-database-revealed-chat-prompts-and-internal-data/
Tomi Engdahl says:
Tom Warren / The Verge:
Microsoft adds DeepSeek’s R1 to Azure AI Foundry and GitHub, and plans to make a distilled, smaller version of R1 available to run locally on Copilot+ PCs soon — Microsoft has moved surprisingly quickly to bring R1 to its Azure customers. — Microsoft has moved surprisingly quickly to bring R1 to its Azure customers.
Microsoft makes DeepSeek’s R1 model available on Azure AI and GitHub
Microsoft has moved surprisingly quickly to bring R1 to its Azure customers.
https://www.theverge.com/news/602162/microsoft-deepseek-r1-model-azure-ai-foundry-github
Matthew Connatser / Tom’s Hardware:
Huawei adds DeepSeek’s R1 to its ModelArts Studio platform, saying the free model is “Ascend-adapted”, referencing its data center GPUs, but offers few details
https://www.tomshardware.com/tech-industry/artificial-intelligence/huawei-adds-deepseek-inference-support-for-its-ascend-ai-gpus
Tomi Engdahl says:
Dario Amodei:
DeepSeek makes US export controls to China even more important, and DeepSeek-V3 is not a unique breakthrough but an expected point on a cost reduction curve — A few weeks ago I made the case for stronger US export controls on chips to China. Since then DeepSeek, a Chinese AI company …
On DeepSeek and Export Controls
https://darioamodei.com/on-deepseek-and-export-controls
New York Times:
Meta executives say DeepSeek’s breakthrough shows that upstarts now have a chance to innovate and compete with AI giants, vindicating its open-source strategy
Meta Engineers See Vindication in DeepSeek’s Apparent Breakthrough
https://www.nytimes.com/2025/01/29/technology/meta-deepseek-ai-open-source.html?unlocked_article_code=1.s04.oaPa.T54dwKwdUhQ3&smid=url-share
The Silicon Valley giant was criticized for giving away its core A.I. technology two years ago for anyone to use. Now that bet is having an impact.
When a small Chinese company called DeepSeek revealed that it had created an A.I. system that could match leading A.I. products made in the United States, the news was greeted in many circles as a warning that China was closing the gap in the global race to build artificial intelligence.
DeepSeek also said it built its new A.I. technology more cost effectively and with fewer hard-to-get computers chips than its American competitors, shocking an industry that had come to believe that bigger and better A.I. would cost billions and billions of dollars.
But A.I. experts inside the tech giant Meta saw DeepSeek’s breakthrough as something more than the arrival of a nimble, new competitor from the other side of the world: It was vindication that an unconventional decision Meta made nearly two years ago was the right call.
In 2023, Meta, in a widely criticized move, gave away its cutting-edge A.I. technology after spending millions to build it. DeepSeek used parts of that technology as well as other A.I. tools freely available on the internet through a software development method called open source.
Meta executives believe DeepSeek’s breakthrough shows that upstarts now have a chance to innovate and compete with the tech giants that have mostly had the A.I. playing field to themselves because A.I. costs so much to build. It was something Meta executives hoped would happen when they gave away their own technology.
“Our open source strategy was validated,” said Ragavan Srinivasan, a Meta vice president, in an interview on Tuesday. “The more people who have access to the technology needed to move things forward faster, the better.”
Meta is also taking a close look at the work done at DeepSeek. Following Meta’s lead, the Chinese company released its technology to the open source tech community as well. Meta has created several “war rooms” where employees are reverse engineering DeepSeek’s technology, according to two people familiar with the effort who spoke on the condition of anonymity.
Before Meta, which owns Facebook, Instagram and WhatsApp, gave away its A.I. tech, the company had been focused on projects like virtual reality. It was caught flat-footed when OpenAI introduced the chatbot ChatGPT in late 2022. Other tech giants like Microsoft, OpenAI’s close partner, and Google were also well ahead in their A.I. efforts.
(The New York Times has sued OpenAI and its partner, Microsoft, claiming copyright infringement of news content related to A.I. systems. The two tech companies have denied the suit’s claims.)
By freely sharing the code that drove its A.I. technology, called Llama, Meta hoped to accelerate the development of its technology and attract others to build on top of it. Meta engineers believed that A.I. experts working collaboratively could make more progress than teams of experts siloed inside companies, as they were at OpenAI and the other tech giants.
Meta could afford to do this. It made money by selling online ads, not A.I. software. By accelerating the development of the A.I. it offered to consumers for free, it could bring more attention to online services like Facebook and Instagram — and sell more ads.
“They were the only major U.S. company to take this approach. And it was easier for them to do this — more defensible,”
Many in Silicon Valley said Meta’s move set a dangerous precedent because the chatbots could help spread disinformation, hate speech and other toxic content. But Meta said that any risks were far outweighed by the benefits of open source. And most A.I. development, they added, had been shared around through open source until ChatGPT made companies leery of showing what they were working on.
Now, if DeepSeek’s work can be replicated — particularly its claim that it was able to build its A.I. more affordably than most had thought possible — that could provide more opportunities for more companies to expand on what Meta did.
“These dynamics are invisible to the U.S. consumer,” said Mr. Nicholson. “But they are hugely important.”
Yann LeCun, an early A.I. pioneer who is Meta’s chief A.I. scientist, said in a post on LinkedIn that people who think the takeaway from DeepSeek’s work should be that China is beating the United States at A.I. development are misreading the situation. “The correct reading is: ‘Open source models are surpassing proprietary ones,’” he said.
Dr. LeCun added that “because their work is published and open source, everyone can profit from it. That is the power of open research.”
By last summer, many Chinese companies had followed Meta’s lead, regularly open sourcing their own work. Those companies included DeepSeek, which was created by a quantitative trading firm called High-Flyer. (On Wednesday, OpenAI claimed that DeepSeek may have improperly harvested its data).
Some Chinese companies offered “fine-tuned” versions of technology open sourced by companies from other countries, like Meta. But others, such as the start-up 01.AI, founded by a well-known investor and technologist named Kai-Fu Lee, used parts of Meta’s code to build more powerful technologies.
Tomi Engdahl says:
Brody Ford / Bloomberg:
IBM reports Q4 revenue up 1% YoY to $17.6B, vs. $17.5B est., bookings for AI consulting and software crossed $5B, up from $3B in Q3; IBM jumps 8%+ after hours — – Software business revenue gained 10%, fueled by Red Hat — IBM gives a forecast for free cash flow that tops estimates
https://www.bloomberg.com/news/articles/2025-01-29/ibm-reports-strong-sales-growth-and-expanded-ai-bookings
Tomi Engdahl says:
Ivan Mehta / TechCrunch:
DeepSeek’s app tops the US Play Store; Appfigures says the DeepSeek app has 1.9M+ downloads on Apple’s App Store and 1.2M+ on Google Play since mid-January 2025
https://techcrunch.com/2025/01/28/deepseek-reaches-no-1-on-us-play-store/
Stefano Bernabei / Reuters:
DeepSeek disappears from Apple’s and Google’s app stores in Italy, a day after Italy’s data protection regulator said it asked DeepSeek about personal data use — The Chinese artifical intelligence application DeepSeek appeared to be unavailable on Wednesday in Apple and Google app stores in Italy on Wednesday.
https://www.reuters.com/technology/deepseek-app-unavailable-apple-google-app-stores-italy-2025-01-29/
Tomi Engdahl says:
Tom Warren / The Verge:
Microsoft adds DeepSeek’s R1 to Azure AI Foundry and GitHub, and plans to make a distilled, smaller version of R1 available to run locally on Copilot+ PCs soon — Microsoft has moved surprisingly quickly to bring R1 to its Azure customers. — Microsoft has moved surprisingly quickly to bring R1 to its Azure customers.
https://www.theverge.com/news/602162/microsoft-deepseek-r1-model-azure-ai-foundry-github
Tomi Engdahl says:
Tom Lee says market sell-off sparked by DeepSeek was an overreaction, as Nvidia suffers its worst day since 2020
https://www.cnbc.com/2025/01/27/tom-lee-calls-market-sell-off-overreaction-nvidia-falls-on-deepseek-news.html
Amazon, Broadcom are buying opportunities on DeepSeek sell-off, Hightower’s Stephanie Link says
https://www.cnbc.com/2025/01/27/amazon-broadcom-are-buys-on-deepseek-sell-off-hightowers-stephanie-link-says.html
Tomi Engdahl says:
Financial Times:NEW
Mistral AI, seen as Europe’s great hope for AI, is losing ground to US rivals and being overshadowed by DeepSeek; sources say its ARR is in the tens of millions
Has Europe’s great hope for AI missed its moment?
Mistral AI was hailed as a potential global leader in the technology. But it has lost ground to US rivals — and now China’s emerging star
https://www.ft.com/content/fa8bad75-dc55-47d9-9eb4-79ac94e54d82?accessToken=zwAGLOZu3EMIkdP6i6113FVH2dOetHmslOVNgg.MEUCIHIBcc_QziJfLrewzPyQebIEbxl3OIYgIKdsBE_VohjpAiEAuWrCZyGOyNba6YaR6LTO-qdXqPokZXVyS7It8d_1MOg&sharetype=gift&token=86b842e6-b986-47b2-a031-66b67f7f4858
True to the strong winds that inspired its name, French start-up Mistral AI took Davos by storm in 2024, having delivered a world-class artificial intelligence model with a fraction of the usual resources.
The Paris-based start-up, less than a year old, was on a high. It was freshly valued at $2bn and had the backing of AI chip leader Nvidia and prominent venture firm Andreessen Horowitz. Mistral’s founding trio of hotshot AI researchers — Guillaume Lample, Timothée Lacroix and chief executive Arthur Mensch — were hailed as the heroes who would finally put Europe at tech’s top table.
Mistral also had the enthusiastic support of French President Emmanuel Macron, who was drawn in by the start-up’s promise of “sovereign” and more “open” AI, proudly independent of US Big Tech.
But a year is a long time in AI. Excitement about Mistral started to cool as it was seen to be struggling to keep up with its larger rivals in the AI race.
Then, this week, came a blast of cold from the east. China’s DeepSeek stunned Silicon Valley by releasing a cutting-edge open-source model with what it claims is a tiny fraction of OpenAI or Meta’s resources and computing power — beating Mistral at its own game.
Tomi Engdahl says:
Wall Street Journal:
US officials and Google researchers: China, Iran, and 18+ other countries use AI tools like Gemini to bolster their cyberattacks against US and global targets
https://www.wsj.com/tech/ai/chinese-and-iranian-hackers-are-using-u-s-ai-products-to-bolster-cyberattacks-ff3c5884?st=SaEfSM&reflink=desktopwebshare_permalink
Chinese and Iranian Hackers Are Using U.S. AI Products to Bolster Cyberattacks
Researchers outline malicious uses of AI after China-built AI platform DeepSeek upends international assumptions about Beijing’s capabilities
Hackers linked to China, Iran and other foreign governments are using new AI technology to bolster their cyberattacks against U.S. and global targets, according to U.S. officials and new security research.
In the past year, dozens of hacking groups in more than 20 countries turned to Google’s Gemini chatbot to assist with malicious code writing, hunts for publicly known cyber vulnerabilities and research into organizations to target for attack, among other tasks, Google’s cyber-threat experts said.
While Western officials and security experts have warned for years about the potential malicious uses of AI, the findings released Wednesday from Google are some of the first to shed light on how exactly foreign adversaries are leveraging generative AI to boost their hacking prowess. This week, the China-built AI platform DeepSeek upended international assumptions about how far along Beijing might be the AI arms race, creating global uncertainty about a technology that could revolutionize work, diplomacy and warfare.
Groups with known ties to China, Iran, Russia and North Korea all used Gemini to support hacking activity, the Google report said. They appeared to treat the platform more as a research assistant than a strategic asset, relying on it for tasks intended to boost productivity rather than to develop fearsome new hacking techniques. All four countries have generally denied U.S. hacking allegations.
“AI is not yet a panacea for threat actors and may actually be a far more important tool for defenders,” said Sandra Joyce, vice president of threat intelligence at Google. “The real impact here is they are gaining some efficiency. They can operate faster and scale up.”
Current and former U.S. officials said they think foreign hacking units are turning to other chatbots as well. Last year, OpenAI also revealed some information about five foreign hacking groups using ChatGPT and said it had disabled the accounts associated with them. That research likewise found that cyberattackers weren’t using ChatGPT for generating significant or novel cyberattacks. A Google spokeswoman said the company terminated accounts linked to malicious activity outlined in its report but declined to disclose how many accounts in total were disrupted.
The company found that a range of sophisticated hacking groups—also known as advanced persistent threats—were using Gemini, but that Chinese and Iranian groups had relied on the tool the most.
More than 20 China-linked groups and at least 10 Iran-linked groups were seen using Gemini, Google said, making them easily the most active countries seeking to use the chatbot.
China was the next most frequent user of Gemini, the report said
“They’re using Gemini to get a leg up in crafting their victim lists and probably improving the effectiveness of the human-directed parts of their operations,” Galante said. She added that large-language models didn’t appear to be “a game changer in terms of the scale of compromises or enabling new tactics or novel operations—but these are still the relatively early days.”
Despite modest uses of generative AI so far, both the U.S. and China see AI technologies as pivotal to future supremacy. The possibility that China’s DeepSeek is rivaling top-tier AI models for a fraction of the cost sent shock waves through Silicon Valley and Washington this week. Unlike Google, DeepSeek’s creators have released their product’s source code, making its misuse harder to track and virtually impossible to prohibit.
DeepSeek’s low cost could have significant national-security implications, too. For years, senior U.S. intelligence officials have warned that China and other adversaries are racing to develop and deploy AI systems to support—and in some cases supplant—their existing military and intelligence objectives.
“America holds the lead in the AI race—but our advantage may not last,” Walker said.
Tomi Engdahl says:
Bloomberg:
Sources: the Trump administration is exploring additional curbs on the sale of chips to China to cover Nvidia’s scaled-down H20, designed for the Chinese market
https://www.bloomberg.com/news/articles/2025-01-29/trump-officials-discuss-tightening-curbs-on-nvidia-china-sales
IT Telkom says:
What are some potential challenges and opportunities that businesses may face as AI continues to evolve? Universitas Telkom
Tomi Engdahl says:
https://www.wiz.io/blog/wiz-research-uncovers-exposed-deepseek-database-leak
Tomi Engdahl says:
https://www.tweaktown.com/news/102887/deepseek-database-contains-chat-history-internal-secrets-for-anyone-to-see/index.html?utm_source=dlvr.it&utm_medium=facebook&fbclid=IwZXh0bgNhZW0CMTEAAR3gIdmPtR_9Hf0KDidjWCQnzSEQ8mRqUTNR7trviD3HZnKhGVX1SV6GlH0_aem_9ULA5kXXIkF_huUyfhYpjg
Tomi Engdahl says:
“In a recent interview with the Chinese media outlet LatePost, Kai-Fu Lee, a veteran entrepreneur and former head of Google China, said that only “front-row players” typically engage in building foundation models such as ChatGPT, as it’s so resource-intensive.”
In 2023, in an interview with Bloomberg, Kai-Fu Lee told the news outlet that his own company, 01.AI, went on a shopping spree and stockpiled “18 months” of Nvidia chips before sanctions were imposed in 2022. This was due to his skepticism that China would be able to produce a quality product.
Currently Nvidia supplies GPUs to China that perform at half of the speed of their top products due to sanctions.
https://www.technologyreview.com/2025/01/24/1110526/china-deepseek-top-ai-despite-sanctions/
DeepSeek’s release of its AI model sent shockwaves through the U.S. stock market most effected was Nvidia whose market value fell by about $590 billion Monday, rose by roughly $260 billion Tuesday and dropped $160 billion Wednesday morning.
As Nvidia senior research manager Jim Fan put it on X: “We are living in a timeline where a non-US company is keeping the original mission of OpenAI alive — truly open, frontier research that empowers all. It makes no sense. The most entertaining outcome is the most likely.”
DeepSeek has a free website and mobile app even for U.S. users with an R1-powered chatbot interface very similar to OpenAI’s ChatGPT.
Tomi Engdahl says:
The brass balls on these guys: OpenAI complains that DeepSeek has been using its data, you know, the copyrighted data it’s been scraping from everywhere
https://www.pcgamer.com/gaming-industry/the-brass-balls-on-these-guys-openai-complains-that-deepseek-has-been-using-its-data-you-know-the-copyrighted-data-its-been-scraping-from-everywhere/