Hugging Face dropped Chat Assistants, the open-source version of OpenAI GPTs.

Here's what stands out:

1. It's completely free to use & share (compared to needing a $20 subscription)

2. You can select from any open-source LLMs to run your Assistant on. Choose from Mixtral, Llama 2, Meta's new Code Llama 70B, and more.

Compared to OpenAI's custom GPTs, it's still missing RAG, web search, actions, and a GPT builder.

This puts it below the custom GPT level for now, but it's all in the roadmap for the open-source competitor.

All around, it's a nice free alternative.

Leak reveals Gemini Ultra coming this week. 3 major Google AI announcements coming on Feb 7th:

1. Bard to be rebranded as Gemini
2. 'Gemini Advanced' (AKA Bard Advanced) tier built on Gemini Ultra launching this week
3. Gemini coming to Canada

Foxconn Chairman said global AI server makers continue to face a severe shortage of AI chips, and even if supplies improve in the 2nd half of 2024, it won’t keep up with demand, noting Foxconn gets more Nvidia GPUs than others because it is the world’s biggest server maker, also making 70% of the world’s AI server-use GPU modules and substrates.

TravelPlanner. A Benchmark for Real-World Planning with Language Agents

TravelPlanner a new planning benchmark that focuses on travel planning, a common real-world planning scenario.

It provides a rich sandbox environment, various tools for accessing nearly four million data records, and 1,225 meticulously curated planning intents and reference plans.

Comprehensive evaluations show that the current language agents are not yet capable of handling such complex planning tasks-even GPT-4 only achieves a success rate of 0.6%.

Language agents struggle to stay on task, use the right tools to collect information, or keep track of multiple constraints.

However, note that the mere possibility for language agents to tackle such a complex problem is in itself non-trivial progress.

TravelPlanner provides a challenging yet meaningful testbed for future language agents.

Amazon unveiled “Rufus".

The AI shopping assistant answers shopping questions, recommends products, and even compares items.

Alibaba releases Qwen 1.5

largest open-source Qwen1.5-72B-Chat, exhibits superior performance, surpassing Claude-2.1, GPT-3.5-Turbo-0613, on both MT-Bench and Alpaca-Eval v2

AI & healthcare brings some of the most fascinating updates.
A recent groundbreaking paper from Stanford and StabilityAI researchers proposes CheXagent.

It's an advanced foundation model specifically designed for interpreting chest X-rays.

CheXagent is a significant leap in AI-driven medical imaging analysis.

It's structured around 3 key components:

- a vision encoder
- a vision-language bridager
- a language decoder

aiming to generate accurate responses from CXR images and associated text.

Its training involved a four-stage process:

- development of a clinical LLM adapted from general-domain models
- training a vision encoder on image-text pairs
- training a bridger to map visual data to language
- instruction tuning for CXR interpretation.

CheXagent significantly outperformed both general and medical-domain Foundation Models (FMs) across a spectrum of clinically relevant tasks.

For example, in view classification and disease identification tasks, CheXagent demonstrated an average accuracy improvement of 97.5% over general-domain FMs and 55.7% over medical FMs.

A human evaluation study involving five radiologists further substantiated the model's capabilities. CheXagent was found to be comparable to physician-generated text in findings summarization, though there remains a gap in findings generation. Moreover, a fairness evaluation revealed disparities in CheXagent's performance across different demographic groups, highlighting the ongoing challenge of bias in AI models used in healthcare.

According to Solscan and OKLink block explorers, the Solana blockchain is suspected of being down, with the last transaction stuck at 17:52 UTC+8.

Upbit announced that due to the interruption, deposits and withdrawals of SOL, GMT, RAY, and ACS will be suspended.

7B Text-to-SQL model outperforms OpenAI GPT-4 Turbo

Defog.AI released a new fine-tuned Code Llama 7B model outperforming the latest GPT-4 & GPT-4 Turbo Models.

The team boosted the 7B performance by leveraging distillation from a fine-tuned 70B Code LLama model!

Models can be commercially used (CC-by-SA-4.0).

OLMo is a new open-source LLM from Allen institute for AI that people are discussing

OLMo stands out by providing the research community with access to "a truly open language model" framework.

It is a holistic toolset for LM research, providing resources for every stage of LM development.

This includes tools for dataset building and analysis, evaluation frameworks, and instruction-tuning methodologies. By offering these tools, OLMo enables more nuanced and comprehensive studies of LMs than previously possible.

The model is presented in two variants: 1B, and 7B scale with a 65B model coming soon.

Weights
Code
Data
Evaluation

Google’s Lookout App released a new feature called ‘Image Question and Answer'.

The AI provides audio responses describing an uploaded image.

A major unlock for helping the blind and partially-sighted better perceive the world.