Sep 28 · AI Daily

People on X always say platforms like BlueSky are full of people who think AI is useless (it's not as extreme as it used to be), but LinkedIn is packed with this outrageous, half-technical garbage advice.

Why aren't they reporting the BLEU, ROUGE, and BERTScore for Astra? 🤔

Good news: Anyone can sign up now 💪

Bad news: We have to temporarily turn off the free tier (for new users only) -- we really want everyone to try jev, but a few bad actors are making it rough for everyone 😢

[Quoting @typesafeai]: Ladies and gentlemen, agents and assistants,

We are beyond thrilled to announce

Jev is back.

Capacity has been scaled up, and signups are officially open!

Ladies and gentlemen, agents, and assistants,

We are beyond thrilled to announce

Jev is back.

Capacity has been scaled up, and signups are now open!

Apollo Chief Economist Torsten Slok believes agentic AI assistants like Muse could soon automatically move household cash into accounts yielding 3.3% to 5.0% instead of leaving it in checking accounts averaging just 0.1%.

I don't think Opus 5.5 could've been made without smoking weed.

After watching this, you'll finally understand PPO, GRPO, and DPO!

Better than expected.

The Honor Choice AI Projector Ultra officially went on sale on Sept 28 at 14:25, with an original price of 2,599 yuan and a national subsidy price of 1,999 yuan. The device features 1,700 CVIA lumens of brightness, uses a self-developed fully enclosed optical engine 2.0 technology, has a native resolution of true 1080P, is powered by a V670 8-core chip with 2GB+64GB of storage, and comes equipped with an AI image calibration system and AI spatial pointing remote control 2.0.

Does anyone know if @huggingface (owned by @NVIDIA) still has access to the MI455X that AMD provided to them under the embargo? Does anyone know if AMD is still continuing to provide exclusive access to next-gen hardware to Hugging Face (owned by NVIDIA)? Will Hugging Face work on getting Nemotron3 Ultra to run on the MI455X?

Lofi Cities debuted on Hacker News as a Show HN, presenting city nightscapes in pixel art with lofi audio generated in real time by the browser. The project earned 104 HN Points, and users can visit loficities.com directly in their browser to try it.

Personal AI agent Muse launched a referral campaign: new users who redeem the referral code S7YZC4 in "Settings" within 48 hours of joining will each get 1 billion Muse tokens. The poster mentioned that the first thing they did after opening Muse was register for a Claude.

When a user searched for the 2010s NBA fan meme "hes never coming over dario," Google's AI Overview mistook it for being abandoned by a man named Dario and gave an empathetic comfort reply, while the actual old tweets and Reddit links wanted were pushed hundreds of pixels down the page. The author questioned why search engines have turned into emotional listeners rather than organizing and presenting information.

Mark Cuban believes humanoid robots will fail within 5-10 years.

Instead, robots will adopt the optimal form factor for their working environment. Homes will be redesigned around dedicated robots, keeping human living spaces separate.

Anthropic CEO Dario Amodei is scheduled to have dinner with President Trump at the White House that evening, marking the first one-on-one meeting between the two. Axios first reported this news, and TechCrunch confirmed it with people familiar with the matter. The two have recently taken opposing stances on AI safety issues. Previously, the Pentagon placed Anthropic on a supply chain risk list over restrictions placed on the use of its technology, and Anthropic is currently litigating the matter.

According to the Financial Times on Sept 26, dark web markets are selling access to AI models from Anthropic, Google, and OpenAI at discounts of up to 97%, compared to top-tier ChatGPT and Claude subscriptions costing $200 per user per month.

Fireworks Research released Ember-1, a specialized model built on Kimi K3 that generates shorter reasoning trajectories, reducing token usage by about 40% while maintaining top-tier quality. Author Elvis Saravia commented that this is a push for token efficiency optimization on the Pareto frontier, noting that the direction remains underexplored as labs rush to ship new models, leaving certain aspects unoptimized.

The SkillGym framework translates human-written skills into 2,756 training environments with code checkers, and fine-tunes on 8,364 successful trajectories.

Chamath: If you ban open-source AI in the US, you might crash the stock market. Businesses would be forced to use models 50-100x more expensive, hammering earnings and valuations. Closed-source model providers would only capture US revenue while losing valuation.

Qualitatively speaking, the gap between open-source and closed-source models is now the widest it has been in a long time. Models at the Fable/Astra level possess agentic capabilities that previous Fable models lacked, for better or worse. No open-source model has crossed this line yet. Once one does, it will be a major leap.

Music startup Thoughtful Things launched a Kickstarter campaign for its first instrument, Engram, a sampler and groovebox running "tiny AI" models locally to warp input audio and generate entirely new hallucinatory timbres.

CodeRabbit hosted the first Jev hackathon at its headquarters, with 160+ developers coding for about 4 hours, and invited @allietheicon to share how to collaborate with new types of AI models like Jev. She discussed the paradigm shift Jev brings to coding and software development, and provided practical tips and examples.

Interesting, this is another version of Paul Graham's essay How to Do Great Work

Given opus 5.5 a simple prompt with the "extra" effort setting

The exact same prompt: "Create a video essay from Paul Graham's essay How to do great work with punchy motion graphics, modern styling and high professionalism"