News from the AI & ML world
DeeperML - #llms
|
Ryan Daws@AI News
//
Anthropic has announced that its AI assistant Claude can now search the web. This enhancement allows Claude to provide users with more up-to-date and relevant responses by expanding its knowledge base beyond its initial training data. It may seem like a minor feature update, but it's not. It is available to paid Claude 3.7 Sonnet users by toggling on "web search" in their profile settings.
This integration emphasizes transparency, as Claude provides direct citations when incorporating information from the web, enabling users to easily fact-check sources. Claude aims to streamline the information-gathering process by processing and delivering relevant sources in a conversational format. Anthropic believes this update will unlock new use cases for Claude across various industries, including sales, finance, research, and shopping. Recommended read:
References :
@the-decoder.com
//
Perplexity AI has launched Deep Research, an AI-powered research tool aimed at competing with OpenAI and Google Gemini. Using DeepSeek-R1, Perplexity is offering comprehensive research reports at a much lower cost than OpenAI, with 500 queries per day for $20 per month compared to OpenAI's $200 per month for only 100 queries. The new service automatically conducts dozens of searches and analyzes hundreds of sources to produce detailed reports in one to two minutes.
Perplexity claims Deep Research performs 8 searches and consults 42 sources to generate a 1,300-word report in under 3 minutes. The company says that Deep Research tool works particularly well for finance, marketing, and technology research. The service is launching first on web browsers, with iOS, Android, and Mac versions planned for later release. Perplexity CEO Aravind Srinivas stated he wants to keep making it faster and cheaper for the interest of humanity. Recommended read:
References :
@Google DeepMind Blog
//
References:
Google DeepMind Blog
, AI News
,
ARC Prize has launched ARC-AGI-2, its toughest AI benchmark yet, accompanied by the announcement of their 2025 competition with $1 million in prizes. ARC-AGI-2 aims to push the limits of general and adaptive AI. As AI progresses beyond narrow tasks to general intelligence, these challenges aim to uncover capability gaps and actively guide innovation. ARC-AGI-2 is designed to be relatively easy for humans, who can solve every task in under two attempts, yet hard or impossible for AI, focusing on areas like symbolic interpretation, compositional reasoning, and contextual rule application.
The benchmark includes datasets with varying visibility and includes the following characteristics: symbolic interpretation, compositional reasoning and contextual rule application. Most existing benchmarks focus on superhuman capabilities, testing advanced, specialised skills. The competition challenges AI developers to attain an 85% accuracy rating on ARC-AGI-2’s private evaluation dataset. Recommended read:
References :
Divya@gbhackers.com
//
References:
gbhackers.com
, Talkback Resources
,
Researchers from Duke University and Carnegie Mellon University have successfully jailbroken several leading AI language models, including OpenAI’s o1/o3, DeepSeek-R1, and Google’s Gemini 2.0 Flash. The team developed a novel attack method called Hijacking Chain-of-Thought (H-CoT), which exploits the reasoning processes of these models to bypass safety mechanisms designed to prevent harmful outputs. This research highlights significant security vulnerabilities in advanced AI systems and raises concerns about their potential misuse.
The researchers introduced the Malicious-Educator benchmark, which utilizes seemingly harmless educational prompts to mask dangerous requests. They found that all tested models failed to consistently recognize these contextual deceptions. For example, DeepSeek-R1 proved particularly susceptible to financial crime queries, providing actionable money laundering steps in a high percentage of test cases. The team has shared mitigation strategies with affected vendors. Recommended read:
References :
@Google DeepMind Blog
//
References:
Google DeepMind Blog
, THE DECODER
Researchers are making strides in understanding how AI models think. Anthropic has developed an "AI microscope" to peek into the internal processes of its Claude model, revealing how it plans ahead, even when generating poetry. This tool provides a limited view of how the AI processes information and reasons through complex tasks. The microscope suggests that Claude uses a language-independent internal representation, a "universal language of thought", for multilingual reasoning.
The team at Google DeepMind introduced JetFormer, a new Transformer designed to directly model raw data. This model, capable of both understanding and generating text and images seamlessly, maximizes the likelihood of raw data without depending on any pre-trained components. Additionally, a comprehensive benchmark called FACTS Grounding has been introduced to evaluate the factuality of large language models (LLMs). This benchmark measures how accurately LLMs ground their responses in provided source material and avoid hallucinations, aiming to improve trust and reliability in AI-generated information. Recommended read:
References :
Matthew S.@IEEE Spectrum
//
References:
IEEE Spectrum
, Composio
Recent research has revealed that AI reasoning models, particularly Large Language Models (LLMs), are prone to overthinking, a phenomenon where these models favor extended internal reasoning over direct interaction with the problem's environment. This overthinking can negatively impact their performance, leading to reduced success rates in resolving issues and increased computational costs. The study highlights a crucial challenge in training AI models: finding the optimal balance between reasoning and efficiency.
The study, conducted by researchers, tasked leading reasoning LLMs with solving problems in benchmark. The results indicated that reasoning models overthought nearly three times as often as their non-reasoning counterparts. Furthermore, the more a model overthought, the fewer problems it successfully resolved. This suggests that while enhanced reasoning capabilities are generally desirable, excessive internal processing can be detrimental, hindering the model's ability to arrive at correct and timely solutions. This raises questions about how to effectively train models to utilize just the right amount of reasoning, avoiding the pitfalls of "analysis paralysis." Recommended read:
References :
@github.com
//
References:
github.com
, LessWrong
Latent Adversarial Training (LAT) has emerged as a promising method for enhancing the safety of Large Language Models (LLMs). A recent study compared LAT to standard Supervised Safety Fine-Tuning (SSFT) and Embedding Space Adversarial Training (AT) and found that LAT encodes refusal behavior in a more distributed way across the model's latent space. This means that instead of relying on a few specific elements, refusal is woven into the model's overall structure, potentially making it more resilient. The study investigated this by generating refusal vectors using each method.
The results indicated that refusal vectors computed from the LAT model were more effective at triggering refusal ablation attacks across multiple models, lowering refusal rates when compared to the other approaches. However, paradoxically, the models trained with LAT maintained the highest refusal rates and were more robust overall against these attacks. This is likely because LAT allows the models to explore a wider range of responses through hidden layer perturbations creating a more comprehensive understanding of refusal. However, the researchers also highlight a potential downside as the more robust encoding of refusal behaviour could be exploited by malicious actors leading to more effective refusal attacks. Recommended read:
References :
Ali Azhar@AIwire
//
References:
AIwire
Microsoft is reportedly developing its own large language models (LLMs), internally called MAI, to compete directly with OpenAI, a company in which Microsoft has invested billions. The move signifies a potential shift in Microsoft's AI strategy, aiming to reduce its dependence on external partners and decrease the costs associated with using OpenAI's models. Microsoft is experimenting with integrating these LLMs into its existing AI products, particularly those based on Microsoft Copilot, as well as Microsoft Teams and Azure cloud services.
Microsoft plans to release the new LLMs as an API for external developers by the end of the year, enabling them to integrate these models into their own applications. This initiative reflects a desire to optimize AI infrastructure and create models tailored specifically for enterprise applications. While Microsoft maintains close ties with OpenAI, the development of in-house LLMs suggests a growing ambition to control its AI destiny and offer more competitive solutions within the rapidly evolving AI landscape. Recommended read:
References :
@www.eweek.com
//
References:
www.eweek.com
OpenAI has been actively advancing its AI capabilities while also focusing on safety and real-world applicability. The company developed SWE-Lancer, a benchmark designed to evaluate how well large language models (LLMs) can perform in software engineering tasks. This test assessed how much money LLMs, including Claude 3.5 Sonnet and GPT-4o, could earn by completing jobs on platforms like Upwork. While the models showed promise, researchers found that they still struggled to solve the majority of tasks, highlighting the challenges of applying AI to complex real-world scenarios.
In addition to practical applications, OpenAI is dedicated to AI safety research. The company uses prompt evaluation techniques to combat potential misuse, specifically focusing on preventing AI from aiding in bio-weapon research. They have also expanded the accessibility of the Operator AI agent to multiple countries, including India, further integrating AI-powered automation into daily tasks. These efforts demonstrate OpenAI's commitment to both innovation and responsible development in the rapidly evolving field of artificial intelligence. Recommended read:
References :
|
BenchmarksBlogsResearch Tools |