AI Unfiltered by Xenoss: Issue #1, February 2025
Welcome to the first edition of AI Unfiltered by Xenoss: a biweekly newsletter that helps make sense of the latest AI news and understand its long-term impact by sharing the thoughts of AI industry leaders.?
The AI landscape is evolving so rapidly that it’s challenging to keep up with what’s happening. It’s even harder to understand what all of it means for the future of the industry and how each innovation or global partnerships change the face of technology.?
We created AI Unfiltered by Xenoss to add more context to headlines and short news descriptions.?
Think of the newsletter as the campfire where you can catch the takes of AI frontrunners on the most significant releases and stories that made headlines.?
Now, let’s dive right into the latest AI buzz.
DeepSeek open-sources its source code
What happened??
In an X post on Friday, DeepSeek AI announced that it would open-source five code repositories containing key infrastructure components of the company’s large-language models.
As of today, three out of five repos are out in the open:?
Why it matters??
DeepSeek says it itself: opens-sourcing significant infrastructure components helps destroy “ivory towers” and spread “pure garage-energy and community-driven innovation.” Until recently, this was not a standard practice in the AI community (OpenAI and Anthropic, the two most prominent players in generative AI, have closed codebases).?
Some models, like Elon Musk’s Grok, are what one could consider “open-source with an asterisk” since xAI openly released the weights but not the data the model was trained on or the training process.?
DeepSeek’s bold commitment to open-source and the incredible hype wave the release generated in the community means that the tide may be turning and getting us a step closer to making open-source a must-have for releasing large language models.?
Community response
Thomas Wolf , co-founder and Chief Science Officer at Hugging Face shared a detailed rundown on DeepSeek's latest released component, Mixture-of-Experts, explaining why DeepSeek's MoE is so impressive and while having access to it is a big deal for the community.
Nathan Lambert , the founder of Ai2 and writer of Interconnects, shared hope (alongside many in the community) that DeepSeek shares the data it used to train its model.
Predictably, a fair share of the community called DeepSeek the "real OpenAI", calling Sam Altman's company out for failing to live up to its name.
Additionally, after the release of DeepSeek caused a nosedive in NVIDIA stocks, people are worried that the release of components that will improve the efficiency of GPU might cause a similar rift.
OpenAI releases GPT 4.5
What happened:
OpenAI released GPT 4.5, a model that is supposed to have "warmer" and more natural speech compared to GPT's former models. The focus on improving GPT's emotional intelligence over crushing technical performance benchmarks sets GPT apart from its competitors, who are targeting operational efficiency instead.
Top comments
Nathan Lands , co-host of The Next Wave Podcast and the author of The AI Newsletter, said that while GPT stays the number-one model for creative work, it feels more dated than Grok 3 and Claude 3.7. and there are no visible reasoning improvements.
Anthropic releases Sonnet 3.7
What happened: With Claude 3.7 Sonnet, Anthropic AI joined the race of “hybrid reasoning models”. The focus on technical capabilities like coding and math has always been Claude’s differentiation from other LLMs, and the new model is even better at STEM tasks.?
Additionally, Anthropic released Claude Code, an active programming assistant that edits code and manages repositories.?
Top comments
Jamie Lord , Solution Architect at CDS UK , explains why the ability to set a token limit for Claude’s reasoning is significant for companies building with the model, and how the company’s decision to switch between snap-quick answers and in-depth reflection is a drastically different move from what its competitors are offering.?
Chris "The Wiz ??" Alexiuk , Deep Learning at 英伟达 , commends Anthropic for trasnparenty reporting the model’s performance across multiple benchmarks and concludes that “everyone’s favorite coding model just got better.”?
xAI released Grok 3
What happened: On February 17, Elon Musk, as the CEO of xAI, unveiled Grok in a live stream. The model uses ten times the GPU of its predecessor, Grok 2, which enables better knowledge, higher coding and instruction-following abilities.?
By performance benchmarks, Grok 3 beat DeepSeek R1 and o3 mini.?
What followed: despite the impressive technical performance, Grok also has a high “spite index”, meaning users do not want to use the LLM because they do not trust its founder Elon Musk. They may not be wrong to think Elon had a hand in censoring the platform’s replies. In multiple occasions, Grok leaked its system prompt preventing it from sharing negative information about Elon Musk and U.S. President Donald Trump.?
Top commentary
Andrej Karpathy, former director of AI at Tesla, a founding member at OpenAI, and now the founder of EurekaLabsAI, was one of the early Grok testers and assessed the model’s Thinking, deep research, and “random LLM gotchas”.?
Read Andrej’s full tweet to know his take on Grok 3.?
领英推荐
Additionally, Theo Browne , YouTuber and CEO at Ping compared Grok 3’s coding skills to o3-mini and Claude Sonnet 3.5 (the test was carried out before the release of Claude 3.7), and Grok lost by a margin.
Alibaba Group has open-sourced Wan 2.1
What happened
Chinese retail giant just released Wan 2.1, an open-source video generation model that outperforms Sora on key benchmarks. It is also 2.5 times faster than OpenAI’s model.?
Wan 2.1 supports text-to-video, image-to-video, and video-to-audio generation. The model can render text in English and Chinese, supporting multi-image referencing, video inpainting, and outpainting. A smaller version of Wan 2.1 can run on consumer hardware.?
Community response
The AI community seems to agree that Wan 2.1. has “ended” Sora.?
This week, Ethan Mollick , Associate Professor at The Wharton School and the author of Co-Intelligence started an interesting debate on what is holding Google back from creating the ultimate Deep Research considering the unparalleled amount of tech and data the search giant has under its belt.?
While Google released Deep Research before OpenAI did, Ethan points out two critical shortcomings.?
The AI community supported the discussion with insightful takes.?
Rohit Krishnan , the founder of Bodo.ai and author of the book “Building God: Demystifying AI for Decision-makers” backed Ethan’s reasoning pointing out that Google’s technological toolset and data availability are unmatched and it is surprising that Deep Research is not creme de la creme.?
Aarush Selvan , Product Manager at Gemini jumped in to correct Ethan’s take on the “dynamic research” capabilities of Deep Research, pointing out that the tool reflects on its findings before issuing new searches and “summarizing the data”.?
In a new Interconnects essay, Nathan Lambert takes a deep dive into Claude 3.7, explaining in much detail why the new model is a solid improvement but not an industry game-changer and helping the broader AI community make sense of the performance data shared by Anthropic.?
by Azeem Azhar
With China releasing state-of-the-art AI models at an unprecedented rate, one may wonder if the US still “has the chops” to maintain its AI supremacy. In his latest write-up, Azeem Azhar ponders on the topic and reflects on how Silicon Valley has reoriented itself around AI technology.?
by Mike Taylor
How would marketing, sales, and content creation change if, instead of going straight for the human target - a client, top-of-the-funnel lead, or a reader, we could practice appealing to their AI prototypes first? Michael Taylor created a simulator of Hacker News to understand how well AI understands virality and whether it can help humans learn a thing or two about it.?
In the light of Grok's release, Bojan Tunguz, Ph.D. , former Senior Systems Software Engineer at 英伟达 , shared his old take on the "roses are red, violets are blue" meme, that Elon Musk appears to have "doubled down on"(Grok 3 uses 200,000 GPU).
We loved this infographic on simple yet powerful Grok 3 prompts that help automate basic tasks and increase your productivity. These are definitely worth making part of the daily routine, and testing in other LLMs.
Xenoss CRO Maria Novikova ?? ?? Possible, Miami shared her take on the significance of Deep Seek’s #OpenSourceWeek for the AI industry. We believe that the move will encourage companies with proprietary codebases to rethink the foundations upon which they structure operations.
That's a wrap
Covering all the significant announcements the AI industry saw in the last two weeks would likely be humanly impossible. We know we haven't touched upon a lot: The AI Summit in Paris, Perplexity releasing a censorship-free version of DeepSeek, Google's AI-enabled assistant for scientists, or Microsoft's first quantum chip.
However, we hope to have offered a full, context-rich image of the news that made it to this edition of AI Unfiltered.
If you want to get more AI news commentary, subscribe to this newsletter, and follow Xenoss on LinkedIn.
Our CRO Maria Novikova also shares weekly AI news roundups, be sure to catch those as well.
What AI news were you most excited about (among those mentioned or others)? Share your picks in the comments.