By the TensorMax editorial team
· Drawing from sources across the AI industry
Today's top story
model release
Perplexity AI has released pplx-embed-v2-late, a pair of ColBERT-style multimodal embedding models, with 0.6B and 9B sizes, designed for fast and cheap queries, and maximum quality, respectively.
Why it matters. The release of pplx-embed-v2-late by Perplexity AI is a significant development in the AI industry, particularly in the field of multimodal embedding models. With two models of 0.6B and 9B sizes, this release has the potential to impact the market, especially considering the 0.6B model's ability to stay close to 8B rivals while using about 340M active parameters for images. The fact that a 9B index can be searched with 0.6B queries, recovering about half the 9B quality gap on text at 0.6B query cost, is also noteworthy, and the MIT license allowing commercial use adds to its strategic stake, with the 9B model achieving a strong 92.4% score on MADQA.
Perplexity AI has released pplx-embed-v2-late, a pair of ColBERT-style multimodal embedding models, with 0.6B and 9B sizes, designed for fast and cheap queries, and maximum quality, respectively. Both models can retrieve text, images, and rendered PDF pages, and share one embedding space. The models are deployable, with the option to host them yourself, and are available on Hugging Face under the MIT license, which allows commercial use. A hosted Perplexity API endpoint is planned but not yet live. The 0.6B model is designed as a lightweight query encoder that can run on edge devices, and both models need sentence-transformers and transformers, with memory figures estimated at 2 bytes per parameter. The models have shown strong results, with the 9B model achieving 92.4% on MADQA, and the 0.6B model staying close to 8B rivals, but there are also caveats, such as storage growth and self-reported scores. The best use cases for these models include visual document search, low-latency search, and agentic RAG over large PDF or web collections. The release of pplx-embed-v2-late is significant, with the potential to impact the market, and its performance, particularly the 9B model's strong score, and the ability to mix sizes, with a 9B index queried by the 0.6B model, make it a notable development in the AI industry.
More from today
model release
Why it matters. OpenAI's release of solutions to 90 of the top 500 open math problems is a significant development, demonstrating the company's capabilities in advanced mathematical research. With an average budget of three hours of Pro-level compute per question, this achievement showcases the potential for AI to drive progress in mathematics. The impact of this release is substantial, with far-reaching implications for various fields that rely on mathematical advancements.
model release
Why it matters. The introduction of Google Cloud's Gemini agent has significant strategic implications, as it provides a unified AI assistant that can act autonomously and complete work across various devices and platforms. With the ability to access the agent through any channel, including command line, Google Workspace, and Slack, Gemini has the potential to revolutionize enterprise work. Notably, the agent's model choice flexibility, which includes options like Gemini Flash and Anthropic PBC Claude models, allows for cost-effective execution of tasks. Furthermore, the 98% drop in per-token prices since 2024 underscores the growing importance of AI in enterprise settings, and Gemini's cost controls, such as smart routing and spend limits, will be crucial in managing AI spend.
model release
Why it matters. OpenAI's release of progress on over 300 math research problems, including 372 math breakthroughs produced by an internal model, has significant implications for the field of mathematics and cryptography. With 372 breakthroughs, the potential for novel attacks on cryptography is substantial, and experts are warning of potential risks to cryptographic protocols and primitives. The fact that OpenAI's model was able to solve approximately 5% of longstanding open mathematical problems in a single 3-hour attempt raises concerns about the potential for AI to break important cryptographic protocols, which could have far-reaching consequences for the security of digital transactions and communication.
model release
Why it matters. The strategic stake of OpenAI's latest mathematical results is significant, as the release of 722 mathematical proofs generated by a powerful, unreleased AI model demonstrates that AI's advances are likely to continue moving across entirely new fields, with mathematics appearing to be the next area of conquest. This development has the potential to disrupt the field of mathematics, with some mathematicians already questioning the implications of these discoveries on graduate work and the training of a new generation of theorists. The fact that one of the results could have warranted an automatic Fields Medal if achieved by a human underscores the significance of this breakthrough, which may have far-reaching consequences for the field.
model release
Why it matters. The launch of Claude Haiku 5.5 by Anthropic marks a significant strategic stake in the AI market, with the model boasting a 75% lower price point than its predecessor Haiku 4.5. This aggressive pricing move is likely to put pressure on competitors, particularly OpenAI, and could potentially disrupt the market. With a price of $0.10 per million input tokens and $0.50 per million output tokens for requests up to 100K tokens, Haiku 5.5 is positioned to capture a significant share of the high-volume, cost-sensitive tasks market.
research paper
$80.2B
Why it matters. Samsung Electronics' forecast of an $80.2 billion profit in Q3, a 782% increase year-on-year, underscores the strategic stake of its thriving chip business and rising demand for AI services. With revenue expected to hit around KRW195 trillion, up almost 127%, the company is poised to surpass the KRW100 trillion milestone for the first time. This significant increase is driven by an ongoing memory chip shortage, fueled by the growing need for AI services, which is expected to continue with emerging AI agents requiring more memory than chatbots.
Catch up quick
-
NVIDIA researchers introduce PivotOPD, a method for teaching multi-turn AI agents to recover from pivotal mistakes
-
Anthropic's internal index shows Claude led 26% of measured model R&D work in August, up from under 1% in February
-
Epoch AI estimates that achieving a fixed score has become 47% cheaper each quarter since 2023, roughly 13-fold cheaper each year
-
Anthropic PBC plans mega-IPO amid concerns over rogue AI risk
-
Major companies such as BNP Paribas, Bradesco, and Merck adopt Google Cloud's Gemini for workflow automation and productivity gains
-
Anthropic plans to go public with a $2T IPO, and has filed its IPO numbers
-
NVIDIA partners with six financial institutions to mobilize over $500B for AI infrastructure
-
Anthropic seeks $2 trillion valuation in upcoming Nasdaq IPO, with research firm New Constructs calling it the 'most ridiculous IPO of 2026'
-
A single attacker used AI-powered hacking tools to breach multiple South Korean banks, stealing large amounts of data, including over 25,000 records from Shinhan Bank.
-
Vitalik Buterin warns that AI could break cryptography, potentially compromising Bitcoin and Ethereum security
-
FCC approves SpaceX's plan to launch 15,000 next-gen satellites for its direct-to-device mobile service
-
Paramount completes purchase of Warner Bros. Discovery, renames combined company Skydance
-
Shortages of DRAM memory and other chips and components in the AI infrastructure build-out
-
China's data center growth, with 89 data centers built or planned in Ulanqab, Inner Mongolia
-
Jim Cramer says AI stocks, including SpaceX, have an advantage in a market split by higher rates
Also on the desk
The Sean Ellis test.Hit reply: if TensorMax disappeared tomorrow, would you miss it? If yes, what specifically?
|
|