Weekly Tech Bytes with ThurAI (Feb 10- Feb 17, 2025)

Weekly analysis of AI & Emerging Tech news from an analyst's point of view.

1️⃣💥𝐇𝐚𝐜𝐤𝐞𝐫𝐬 𝐮𝐩𝐥𝐨𝐚𝐝𝐞𝐝 𝐦𝐚𝐥𝐢𝐜𝐢𝐨𝐮𝐬 𝐀𝐈 𝐌𝐨𝐝𝐞𝐥𝐬 𝐨𝐧 𝐇𝐮𝐠𝐠𝐢𝐧𝐠𝐟𝐚𝐜𝐞.💥

Details:

These models utilized a novel technique to successfully evade security detection through "corrupted" pickle files. This malicious code mainly consists of a reverse shell that can connect to a hard-coded IP address, enabling hackers to control the system remotely. This attack method using pickle files is known as nullifAI, and it aims to bypass existing security measures. Hugging Face claims to have updated its security scanning tools to fix related vulnerabilities in the future.

Analysis: 

This news while went unnoticed by many, is a pretty big deal. This just proved the models can be corrupt, poisoned, or full of malware and can be easily released in the wild. Most of them will also have names sounding similar to the model from bigger model producers, such as OpenAI or DeepSeek, and claim to be fine-tuned for a certain purpose. Don't trust any model you see in a model catalog even if it is listed on reputable sites like Huggingface. While this particular method of exploitation is taken care of now, it is only a matter of time before others pop up.

2️⃣𝐀𝐧𝐭𝐡𝐫𝐨𝐩𝐢𝐜'𝐬 𝐫𝐞𝐜𝐞𝐧𝐭 𝐀𝐈 𝐬𝐭𝐮𝐝𝐲 𝐬𝐡𝐨𝐰𝐬 𝐚 𝐬𝐥𝐨𝐰 𝐀𝐈 𝐚𝐝𝐨𝐩𝐭𝐢𝐨𝐧

Details:

• According to Anthropic's recent AI study, the adoption of artificial intelligence is currently happening at a slow pace, with most users primarily experimenting with AI tools rather than fully integrating them into their workflows, indicating a significant gap between trial and widespread adoption.
• Most jobs only use AI for a small fraction of their tasks, with very few occupations relying heavily on AI for the majority of their work.
• Computer-related professions, like software engineers, show the most significant level of AI usage compared to other industries.
• Users are primarily using AI to augment their existing work rather than completely replace human tasks.
• The study found that mid-range salary jobs, like computer programmers, are more likely to use AI compared to low-wage or very high-paying positions.

Analysis: 

Some of the report findings are in line with what we see in the field. Software development (coding, testing, etc.) leads the pack along with technical writing tasks (writing blogs, documentation, emails, autoresponders, etc.). But what is surprising is the impact and penetration is very low. Still, a majority use it toward work augmentation (57%) vs full automation so AI can perform the tasks without a human involved (43%). The most surprising fact is that AI use is more prevalent for tasks associated with mid-to-high-wage occupations like computer programmers and data scientists, but is lower for both the lowest- and highest-paid roles. This likely reflects both the limits of current AI capabilities, as well as practical barriers to using the technology. The bottom line, AI is still in the experimentation phase with sporadic adoption. Moving it to production faces a lot of hassle including ethical, trustworthy, safety, and security among others.

🔗Link to the full research report - https://assets.anthropic.com/m/2e23255f1e84ca97/original/Economic_Tasks_AI_Paper.pdf

https://www.axios.com/2025/02/10/anthropic-economic-index-ai-use-data

3️⃣𝗛𝘂𝗴𝗴𝗶𝗻𝗴 𝗙𝗮𝗰𝗲 𝗟𝗮𝘂𝗻𝗰𝗵𝗲𝘀 𝗔𝗴𝗲𝗻𝘁 𝗟𝗲𝗮𝗱𝗲𝗿𝗯𝗼𝗮𝗿𝗱

Details:

Hugging Face just launched an agent leaderboard to track the top LLMs that are used for AI agents. Both open-source and closed-source models were tested for tool selection accuracy, multi-tool orchestration, edge case handling, and real-world performance. A total of 17 models (12 closed source and 5 open source) were evaluated in this process.

Analysis: 

The results are a bit of a surprise to me. Performance-wise, Google Gemini takes home 3 of the top 5 spots including the #1 spot. Meta's Llama didn't even make the top 10. What is shocking is how cheap Google Gemini model usage is. In general, the Gemini model costs equal to or cheaper than some of the open-source models. However, when it comes to performance/cost, Google models clear the board with a clean sweep.

🔗The leader board can be seen here - https://huggingface.co/spaces/galileo-ai/agent-leaderboard

🔗A full blog explaining this process can be seen here - https://www.galileo.ai/blog/agent-leaderboard

4️⃣𝐎𝐩𝐞𝐧𝐀𝐈 𝐚𝐧𝐧𝐨𝐮𝐧𝐜𝐞𝐝 𝐚 𝐦𝐚𝐣𝐨𝐫 𝐫𝐨𝐚𝐝𝐦𝐚𝐩 𝐮𝐩𝐝𝐚𝐭𝐞 𝐭𝐨 𝐬𝐢𝐦𝐩𝐥𝐢𝐟𝐲 𝐀𝐈 𝐞𝐜𝐨𝐬𝐲𝐬𝐭𝐞𝐦!

Details:

• OpenAI just announced plans to merge its fragmented AI ecosystem — combining reasoning models, GPT variants, and advanced tools — into a single intelligence platform "that just works"
• GPT-4.5 (Orion) will be released in the next few weeks. This will be the last standalone model before all models get integrated into one.
• OpenAI will unify all its models – GPT-series and o-series will merge into a single, intelligent system.
• GPT-5 will be a fully unified model. Advanced reasoning (o-series), deep research capabilities, voice, canvas, and search are all integrated into one AI.
• Model selection done by AI - choosing between a reasoning model vs basic LLM

Analysis:

It is about time that all providers did this. To begin with, the AI provider market is fragmented with many AI model providers with varying cost, and performance ratios making it hard to figure out which model to use and when. By offering model consolidation, not only does it make it easier to decide the model choice but also the dynamic model switching can be done based on the usecase. Hopefully, all model providers will take this route. I have always wondered this - is AI not intelligent enough to pick the right model for you 😂

In other news,

• AI.com is now owned by Deepseel’s parent company. The website directs users to chat.deepseek.com.
• The UK and the US are the only major countries that refused to sign the international AI declaration at the AI action summit in Paris. The statement signed by 61 countries, including France, Canada, Germany, Japan, China, India, the African Union, and the EU, pledges an “open, inclusive and ethical” approach to AI. The full text of that statement can be seen here - https://www.elysee.fr/en/emmanuel-macron/2025/02/11/statement-on-inclusive-and-sustainable-artificial-intelligence-for-people-and-the-planet
• Tencent has launched an AI contract drafting feature that can easily generate legal documents. This feature is powered by two major AI models, Hunyuan and DeepSeek, aimed at helping users generate legal documents with one click. They also offer “AI Contract Review” and “AI Contract Management.”
• At the same AI summit, French President Emmanuel Macron unveiled €109 billion in private-sector commitments to advance AI research and infrastructure. EU chairman Ursula von der Leyen also announced a €200 billion investment across the EU for “AI-related opportunities,” including AI gigafactories.
• Apple is reportedly exploring humanoid and non-humanoid robots with a potential mass production in 2028. The biggest challenges are pricing and reliability according to analyst Ming-Chi Kuo.
• AI takes center stage at the Super Bowl. OpenAI, Google, Meta, Salesforce’s Agentforce, GoDaddyls Airo, and Cirkul all had splashy AI ads during the Superbowl. Apparently, the advertisement cost OpenAI close to 14 million. https://techcrunch.com/2025/02/10/ai-driven-ads-take-the-field-during-2025-super-bowl/
• OpenAI is finalizing its first self-developed chip, to avoid reliance on NVIDIA GPUs. TSMC, as the sole producer of the chip, is expected to do the test production phase soon with mass production of the chip expected to begin in mid-2026.
• Per Bloomberg, Meta officially launched a layoff plan primarily targeting "underperforming" employees. The layoffs will affect approximately 3,600 employees, accounting for 5% of Meta's total workforce. US-based employees will receive severance pay including 16 weeks of salary, along with an additional two weeks of pay based on their length of service.
• ByteDance Launches OmniHuman-1: Turning a Photo into a Talking, Lively Virtual Human.
• US courts ruled in favor of Thomson Reuters in a copyright case involving artificial intelligence (AI) technology. Thomson Reuters accused certain AI companies used its copyright-protected data without authorization to train their models and generate similar content. The court determined that Thomson Reuters' data is protected by copyright, and any unauthorized use of this data constitutes infringement.
• At the recent Paris AI Action Summit, OpenAI CEO Sam Altman revealed that OpenAI is willing to engage in deep cooperation with China in the field of artificial intelligence. He also praised the performance of DeepSeek and said their model can compete with OpenAI's ChatGPT. He particularly praised DeepSeek's capabilities in demonstrating thought processes. This is a bit interesting as OpenAI publicly accused DeepSeek of distilling knowledge from their models.
• Christie's Auction House to Hold Its First Auction Featuring Artificial Intelligence Art. The exhibition will take place at Christie's Rockefeller Center gallery in New York, starting from February 20. A highlight of the event will be a live painting robot demonstration.

#GenerativeAI #GenAI #AI #LLM #OpenAI