Wow, what a crazy AI week it has been! Among the top news are OpenAI who just announced free (!) fine-tuning of GPT-4o until September 23, Ideogram 2.0 image generator with amazing text rendering, Meshy-4 the amazing image-to-3D model generator, and Lumalabs Dream Machine 1.5. On top of that we have an AI powered mosquito hunter with built-in laser, roomier robottaxis from Waymo and improved music recommendations using transformers.
THIS WEEK’S NEWS:
- Ideogram Launches version 2.0 with State-of-the-art Text Rendering
- Lumalabs Launches Dream Machine 1.5
- OpenAI Adds Fine-Tuning to GPT-4o
- Meshy-4 Brings Sci-Fi Level AI to 3D Modeling and Design
- Bzigo Launches Iris – a Laser Powered AI Mosquito Detector
- Waymo Announces Roomier and Less Expensive Robottaxis
- OpenAI Signs Deal With Condé Nast
- Google Research Enhances Music Recommendations with Transformers
- California Weakens AI Safety Bill
Ideogram Launches version 2.0 with State-of-the-art Text Rendering
https://twitter.com/ideogram_ai/status/1826277550798278804
The News:
- Ideogram just released version 2.0 of its advanced text-to-image model with major upgrades and new features, including five new image styles, an iOS app, a beta API, and over 1 billion public Ideogram images.
- According to Ideogram, human evaluations consistently rate Ideogram 2.0 as significantly better than Flux Pro and DALL·E 3.
- The updated “Design style” significantly improves text rendering in Ideogram 2.0, and makes it possible to create “premium graphic designs including greeting cards, t-shirt designs, posters, illustrations with longer and more accurate text.”
My take: The cartoon style images generated by Ideogram 2.0 look absolutely stunning, and I can see this new model becoming very popular among YouTube thumbnail generators. The development of image generators is going so quickly now that it’s almost impossible to keep up with them all, but if you need to generate AI-images with text you should definitely check this one out.
Lumalabs Launches Dream Machine 1.5
https://www.tomsguide.com/ai/ai-image-video/luma-drops-dream-machine-15-heres-whats-new
The News:
- Lumalabs has updated their Dream Machine to version 1.5, offering better realism, motion following, and prompt understanding.
- Dream Machine 1.5 excels in smooth motion, cinematography, and dramatic shots, turning static concepts into dynamic stories, but lags in morphing, movement, and text.
My take: I wasn’t too impressed with Dream Machine when version 1.0 launched in June, but version 1.5 looks much better. Definitely give this one a try if you have the need for short 5-second AI generated clips.
OpenAI Adds Fine-Tuning to GPT-4o
https://openai.com/index/gpt-4o-fine-tuning
The News:
- OpenAI has launched free fine-tuning (up to 1 million tokens per day through September 23) for GPT-4o, allowing developers to customize the model for higher performance and accuracy.
- Fine-tuning is available on all paid usage tiers with training costs of $25 per million tokens, but it is completely free until September 23.
- OpenAI suggests that developers should see strong results from fine-tuning with only a few dozen training examples.
My take: Wow, this is probably the greatest news from OpenAI this year. Expect an explosion of highly skilled niche LLMs based on GPT-4o in the coming weeks, where each one of them surpasses every other LLM in a specific area.
Meshy-4 Brings Sci-Fi Level AI to 3D Modeling and Design
https://venturebeat.com/ai/meshy-4-brings-sci-fi-level-ai-to-3d-modeling-and-design/
The News:
- Meshy, a startup in the AI design space, have released Meshy-4, its latest AI-powered 3D modeling tool. The new version offers improved mesh geometry and a redesigned workflow, aiming to change how designers and developers create virtual environments.
My take: The way Meshy can go from a single photo to a complex and UV-mapped 3D Model is nothing short of amazing. If you are working with 3D, have been working with 3D, or plan to work with 3D, do yourself a favor and check their launch video of Meshy-4 below. It’s absolutely amazing.
Read more:
Bzigo Launches Iris – a Laser Powered AI Mosquito Detector
https://www.foxnews.com/tech/ai-technology-can-help-you-win-battle-over-mosquitoes
The News:
- Bzigo Iris is a small box that uses infrared vision to track mosquitos and a small laser pointer to show you where they are hiding.
- The Iris itself does not kill the mosquitos, you have to kill them yourself, but it does track them with very high precision.
My take: Pretty cool little app powered box for $399. If you live in a country where mosquitos are a problem, the Iris could be exactly what you are looking for.
Waymo Announces Roomier and Less Expensive Robottaxis
https://www.cnbc.com/2024/08/19/waymo-generation-6-robotaxi-geely-zeekr.html
The News:
- Alphabet-owned Waymo revealed details about its newest “generation 6” self-driving technology. Its new driverless tech, integrated into Geely Zeekr electric vehicles, should be able to handle a wider array of weather conditions while at the same time requiring less cameras and sensors.
- In part to lower the cost of each robotaxi, Waymo said it has reduced the number of cameras on board from 29 to 13 and lidar sensors from five to four.
- On August 20, Waymo also announced that they have surpassed 100k paid trips per week.
My take: Going from 29 down to 13 cameras is a great accomplishment, and with version 6 of their software stack it seems nothing can stop Waymo from growing rapidly in the coming years. Already at over 100.000 paid fully automatic trips per week, this is the future of taxi and having larger vehicles with more space is definitely a step in the right direction.
Read more:
OpenAI Signs Deal With Condé Nast
https://www.axios.com/2024/08/20/openai-conde-nast-deal-chatgpt-search
The News:
- OpenAI has signed a multiyear licensing deal with Condé Nast, the companies announced last week.
- Condé Nast is home to some of the world’s biggest tech, lifestyle and culture brands, including Vogue, The New Yorker, Bon Appétit, Vanity Fair and Wired.
- The arrangement gives OpenAI license to display content from Condé Nast brands within OpenAI’s products, including ChatGPT and its SearchGPT prototype.
My take: Where previous versions of ChatGPT was trained on data from dubious sources, OpenAI is now signing deals with most top publishers on a weekly basis. It will be interesting to see if other companies will follow suit, or if OpenAI has managed exclusivity in their deals. If this is the case, companies such as Anthropic that was recently sued for copyright infringement will have to look deep for other source materials to train their models.
Read more:
Google Research Enhances Music Recommendations with Transformers
https://research.google/blog/transformers-in-music-recommendation
The News:
- Google has published new research for enhanced YouTube music recommendations, using Transformers.
- The new system takes into account the user’s current context (e.g., working out at the gym) when making recommendations, rather than solely relying on past preferences.
- By adapting the Transformer architecture, typically used in language models, Google’s system can process sequences of user actions to make more nuanced recommendations.
- The new Transformer-based component works in conjunction with existing ranking models, creating a hybrid system that combines the strengths of both approaches.
My take: I have really tried both Apple Music and Spotify’s song recommendations, but since they don’t consider the context (if I am running, training in the gym, relaxing) most of the recommendations are not what I am looking for. Integrating a transformer architecture that considers context could be a solution for that, and I am very much looking forward to this solution in the coming months.
California Weakens AI Safety Bill
The News:
- Earlier this year California announced proposal SB 1047 to prevent ”AI disasters”. The bill faced significant opposition from many parties in Silicon Valley, and last week the Californian lawmakers added in several amendments, suggested by companies such as Anthropic.
- In short, the bill no longer allows California’s attorney general to sue AI companies for negligent safety practices before a catastrophic event occurs. AI labs are now only required to submit public “statements” about their safety practices vs certifications “under penalty of perjury”. Likewise, developers must now provide “reasonable care” vs “reasonable assurance” that AI models do not pose significant risks.
- The bill is headed to California’s Assembly floor for a final vote
My take: It’s positive to see that the Californian government is listening to the industry when it comes to AI safety, especially when it comes to balancing innovation with legislation. It will be interesting to follow how this bill is used in practice next year.