OpenAI has come up w/ a framework of 5 levels to track progress twd AGI, and think they're currently near level 2 ("Reasoners")

At recent all-hands, leadership also did a research demo of GPT4 w/ improved reasoning

Self-employed AI coming soon.

Graphcore has been acquired by SoftBank. Masayoshi Son laid out his grand vision for artificial superintelligence last month, and it now seems clear Graphcore will be a part of that plan.

Robotics is having its ChatGPT moment. Here's a market map of companies

Like how ChatGPT works by predicting the next word or pixel, robots use foundational models to predict the next action.

Patronus AI announced the release of 'Lynx'. It's a new open-source hallucination detection model.

- Beats GPT-4o on hallucination tasks
- Open source, open weights, open data
- Excels in real-world domains like medicine and finance.

The startup claims it outperforms existing solutions such as GPT-4, Claude-3-Sonnet, and other models used as judges in closed and open-source settings.

Hf.

Paper.

A really nice read on the wider set of mathematical tools and approaches that can help model data with perhaps less standard topological or algebraic properties

Commure announced Strongline Copilot - a wearable AI device for nurses, physicians, and healthcare administrators to interact with their EMR, automatically generate SOAP notes, and save valuable time in-clinic.

It's compatible with Strongline platform.

New Google DeepMind scalable oversight paper: “How do you supervise and provide feedback to superhuman AI systems?”

Previous work has proposed protocols like debate (where two AIs argue and a judge decides) and consultancy (where one AI tries to convince a judge) as potential solutions.

This study aims to evaluate these more thoroughly across a wider range of tasks and models than previous work, with Gemini 1.5 used as a debater or consultant, and smaller models (like Gemma7B and Gemini Pro 1.0) are used for judges.

In short, debate generally outperforms consultancy across all tasks in terms of judge accuracy: in open consultancy, judges tend to agree more with the consultant's choice (which may be incorrect).

In open debate, judges are less likely to be convinced when the protagonist chooses incorrectly. Note that for closed tasks (QA w/o the need for additional source material), the results are mixed, with small or no advantage for debate.

So in short: debate tends to work best, but effectiveness varies depending on the task type and specific setup.

Woah qwen2-500M trained on 12 trillion tokens… this has to be the most tokens for a model this size

Alibaba released a comprehensive suite of foundational and instruction-tuned language models, encompassing a parameter range from 0.5 to 72 billion, featuring dense models and a Mixture-of-Experts model.

Qwen2 surpasses most prior open-weight models, including its predecessor Qwen1.5, and exhibits competitive performance relative to proprietary models across diverse benchmarks on language understanding, generation, multilingual proficiency, coding, mathematics, and reasoning.

Google offered a group of European Union-based cloud firms a package worth about €470 million ($512 million) in a failed attempt to derail their antitrust settlement with Microsoft .

Mistral released their first Mamba Model!

Codestral Mamba 7B is a Code LLM based on the Mamba2 architecture. Released under Apache 2.0 and achieves 75% on HumanEval for Python Coding.

HuggingFace.

They also released a Math fine-tuning base on Mistral 7B that achieves 56.6% on MATH and 63.47% on MMLU.

Model.

⚡️ OpenAI co-founder Andrej Karpathy launches AI education school Eureka Labs.

He left OpenAI in February.

According to Andrei on X, Eureka Labs is a new type of school based on AI.

How can we get closer to the ideal experience of learning something new?

Given the recent progress in generative AI, this learning experience seems quite acceptable. The teacher still develops course materials, but they are supported, used and scaled with the help of an AI teacher's assistant, which is optimized to help students in their learning.

This Teacher + AI symbiosis can conduct an entire course program on a common platform.

The first product will be an LLM101n AI course. This is a course for undergraduates that helps train their own AI, very similar to a scaled-down version of the AI teacher's assistant itself. Course materials will be available online, but there are also plans to run both online and offline course groups.

GitHub.

First text2protein AI model, compressing billions of years of life.

800+ novel, functional and foldable proteins are discovered by researchers.

Repo of 800+ curated prompt+results.

Deloitte latest report: Platform business models – #PrivateEquity opportunities in the life sciences and health care sector identifies plays for healthtech and medtech portfolio companies to transition to platform businesses and offers examples of how players are doing this in the market today.

OpenAI trained advanced language models to generate text that weaker models can easily verify, and found it also made these texts easier for human evaluation.

This research could help AI systems be more verifiable and trustworthy in the real world.

Meta said: “We will release a multimodal Llama model over the coming months, but not in the EU due to the unpredictable nature of the European regulatory environment”.

Unless the EU changes course here's what won't be coming to Europe:

- Apple Intelligence
- Agent Memory, all agents
- Llama 4, and beyond

Everybody is talking about ColPali, a new retrieval model architecture that uses vision language models to directly embed page images, without relying on complex text extraction pipelines.

Combined with a late interaction matching mechanism, ColPali largely outperforms modern document retrieval pipelines while being drastically faster and end-to-end trainable.

Menlo Ventures launched the $100M Anthology Fund, an Anthropic partnership to fund Seed and Series As of the next generation of AI startups around the world

Startups will get:

— $25,000 Anthropic credits
— Access to Anthropic's AI models
— quarterly deep dives with the Anthropic team
— biannual demo days hosted by Anthropic CPO Mike Krieger and Cofounder Daniela Amodei
— credits from Menlo Ventures company

Here's OpenAI’s new model:'GPT-4o mini'.

The company called the new release “the most capable and cost-efficient small model available today,” and it plans to integrate image, video and audio into it later.

The mini AI model is an offshoot of GPT-4o, OpenAI’s fastest and most powerful model yet, which it launched in May during a livestreamed event with executives. The o in GPT-4o stands for omni, and GPT-4o has improved audio, video and text capabilities, with the ability to handle 50 different languages with improved speed and quality, according to the company.

Check out OpenAI’s new model GPT-4o mini : 82% MMLU at 60 cents per 1M output tokens!

OpenAI has talked to Broadcom about developing new AI chip

OpenAI has been hiring former members of a Google unit that produces Google’s AI chip, the tensor processing unit, and has sought to develop an AI server chip.

OpenAI has been talking to chip designers including Broadcom about working on the chip.

The team has discussed how the eventual chip could help the new venture Altman has envisioned, which aims to increase the amount of computing power for AI developers such as OpenAI.