AI Roundup 087: DevDay SF
October 4, 2024.
Workshops
I’m very excited to announce our next workshop from Sairam Sundaresan of
! Sairam has worked in AI/ML for over a decade, and will be going through how the architecture of GPTs works, and how you can make OpenAI’s GPTs work for you.And in case you missed it, I’ve published last month’s Prompt Engineering workshop for paid subscribers:
DevDay SF
This week brought the first of three DevDays from OpenAI, with a slew of new updates, APIs, and product releases.
Between the lines:
The announcements include a Realtime voice API, vision fine-tuning, prompt caching, and model distillation (plus a new Canvas launch).
The new releases both bring ChatGPT closer to parity with features from Deepmind and Anthropic, as well as offer new customizations for enterprises.
Meanwhile, the company has also completed its record-breaking $6.5B fundraise (plus $4B in revolving credit), as it navigates a tricky path towards for-profit status.
Elsewhere in frontier models:
Google released Gemini 1.5 Flash-8B, a smaller and faster variant with lower pricing, higher rate limits, and improved latency on small prompts.
Nvidia unveiled its LLM family NVLM 1.0, capable of handling vision and language tasks while enhancing text-only capabilities, led by the 72B-parameter NVLM-D-72B.
And sources report Google teams have made progress on AI reasoning models, similar to OpenAI's efforts with o1-mini and o1-preview.
The AIs are getting ads
Google is rolling out ads in AI Overviews, though only in the US, and only for questions with a "commercial angle."
Why it matters:
With so much AI development being subsidized by tech giants and tech investors, many have wondered what business models will emerge besides subscriptions.
Specifically for Google, it's been unclear whether there's an ad-based model that will still work (and still be as lucrative) with generative AI.
But Google isn't the only one trying out different monetization strategies - Perplexity is still expected to launch "sponsored questions" later this year.
Elsewhere in Google:
Gemini-powered contextual Smart Replies in Gmail for paid users.
An update for Lens that lets users take a video and use their voice to search, showing an AI Overview and search results.
AI-organized Search results in the US on mobile, aggregating content from various sources.
An expansion Gemini Live's language support to French, German, Portuguese, Hindi, and Spanish.
And DeepMind is developing an AI research assistant for scientists, while BioNTech's InstaDeep unveils Laila for experimental biology.
Elsewhere in the FAANG-free-for-all Microsoft:
The Windows 11 update adds generative AI-based features to Paint and Photos for Copilot+ PCs.
Bing Generative Search begins rolling out to all US users, underpinned by a mix of AI models.
Microsoft plans to pay outlets for content in Copilot Daily, a new assistant summarizing current events.
Copilot Vision, a new feature which analyzes text and images on webpages, along with other new Copilot features.
And Microsoft details how it has overhauled its controversial AI-powered Recall feature, making it opt-in and allowing users to uninstall.
The noes have it
With one day left to decide, Governor Newsom vetoed SB 1047, a bill we've previously discussed.
Where we're headed:
The governor didn't veto the bill because it was too onerous for AI companies - rather, he felt it only penalized the largest models, while smaller models in high-risk contexts wouldn't be covered.
The bill's sponsors have said they plan to continue the fight, and are open to reintroducing a modified version of the bill in a future session.
But future bills may also face resistance from the judicial system - a federal judge has already blocked one of California's new election deepfake laws, passed less than two weeks ago.
Elsewhere in AI anxiety:
A network of AI "nudify" sites has been revealed to be a front for notorious Russian hackers.
Udemy, a massive e-learning platform, has already closed its Gen AI opt-out window for teachers.
And Neo-Nazis are using AI to create English versions of Hitler's speeches and rebrand him as "misunderstood" on social media platforms.
Things happen
Accenture plans to train 30,000 employees on Nvidia's AI tech and launch a new Nvidia Business Group. OpenAI reportedly asks investors for an exclusive funding arrangement to maintain its generative AI lead. Character.AI pivots to focus on chatbots after Google poaches its founders. Microsoft AI CEO Mustafa Suleyman discusses progress on AI agents that can transact on users' behalf. SAP CEO Christian Klein argues against AI regulation in the EU and discusses the company's €2B annual AI investment. Researchers explore looking inside Llama 3 to understand and steer AI models. Calliope Neworks launches License to Scrape program for YouTubers to strike AI training data agreements. YC-backed PearAI faces criticism for forking open-source AI code and applying a closed license. Google NotebookLM's Audio Overview feature generates surprisingly effective AI-hosted podcasts. AI chp startup Cerebras Systems files for IPO, reporting $66.6M net loss on $136.4M in sales for H1 2024. Inside Elon Musk's AI party at OpenAI's old headquarters. NYU profesor Aswath Damodaran discusses an AI bot trained on his work and why AI will empower generalists. UK CMA clears Amazon's $4B Anthropic deal as it doesn't meet investigation thresholds. MIT spinoff Liquid AI debuts non-transformer AI models claiming state-of-the-art performance. Stripe data shows AI startups reach $1M revenue faster than previous SaaS startups. Anthropic hires OpenAI co-founder Durk Kingma. OpenAI faces intense competition and massive burn rate as AI models become "more of a commodity." China reportedly discourages local companies from purchasing Nvidia's H20 chips. Stop using generative AI for hero images. Study suggests AI can outperform human CEOs in most tasks.
