They are literally queueing out of the host hotel front door for this morning's panel on AI and small newsrooms in Perugia #ijf24

Anyone interested in the topic of this panel should check out the detailed report on the AI in Journalism Challenge:

opensocietyfoundations.org/publications/o…

I'm really looking forward to the International Journalism Festival in Perugia! Lots of AI talk this year!

journalism festival
#ijf24

In the 2nd half of 2023 I was proud to lead the AI in Journalism Challenge - an AI accelerator programme for 12 newsrooms from across the globe, funded by Open Society Foundations.

This report described the programme, projects, lessons, recommendations & next steps.

opensocietyfoundations.org/publications/o…

“Listening as a service: What you are selling is not content - but understanding the community you serve.”

One of David Caswell’s great points during our talk at #NAMS24

(The podcast is in Danish, but listen to the interview with David from 7:30 - link below)

.Zach Seward's thoughts on creating structure with generative AI.

Structured journalism isn't superseded by generative AI - it's enabled by generative AI.

In fact, journalism might depend on structure when media become a soup of language in an LLM world.

zachseward.com/creating-struc…

This was a fun discussion with Peder Hammerskov at the Nordic AI in Media Summit!

The interview (in English) starts at 7:30, but my Danish clone puts in an appearance earlier.

Also on Spotify.

#NAMS24

podcasts.apple.com/dk/podcast/ep-…

What a great few days at the Nordic AI in Media Summit!

This year emphasized the transition from exploration to application and my panel with Lily, Sultan Suleimanov & Jody Doherty-Cove was part of that.

It's thrilling to watch this entire field develop and mature so quickly! #NAMS24

I just presented: Reality Check, reviewing and previewing the AI shift at the Nordic AI in Media Summit. A short recap of what happened since early 2023 and a look at what is still to come. Slides can be found here: linkedin.com/posts/ezra-eem…?

Canada has announced new AI investment ahead of their budget next week:

$2 billion for compute
$300 million for new applications
$50 million for a new AI safety institute
$50 million for 'workers impacted by AI'
$5 million for regulation enforcement

pm.gc.ca/en/news/news-r…

I'm excited to be attending the 2nd Nordic AI in Media Summit (NAMS) in Copenhagen next week.

I'll be moderating a panel with Lily, Sultan Suleimanov & Jody Doherty-Cove, titled 'Implementing AI in Newsrooms: Lessons from small, medium and large publishers'.

#NAMS24

nordicaijournalism.com/nams2024

The New York Times may be moving to AI audio versions of all of their articles.

This is similar to the work that the NYT's Mukul Devichand led in the BBC's Voice & AI team, which the BBC unfortunately shut down just before AI took off!

axios.com/2024/04/02/exc…

New fine-tuning stuff from OpenAI, including a side-by-side 'Comparative Playground' for comparing fine-tuning with prompts, sharing fine-tuning with 3rd parties, new validation metrics, more hyperparameters and Improvements to the fine-Tuning Dashboard.

openai.com/blog/introduci…

Yet another new GPT-4 level LLM with a humungous context window...

This is like search in the late 90's, but without any obvious 'winner take all' opportunity for sustained competitive advantage.

I had assumed this was April Fools content, but apparently not...

theverge.com/2024/4/2/24118…

The assumption-busting that goes on here is useful as a metaphor for the risk from fast-takeoff AGI.

Most AI risk fits well within existing economic, cultural and technical expectations, but some doesn't.

We should take AI safety seriously.

Playing with anything-to-anything AI models gives me a sense of what next-gen AI might look like.

It may look like a consolidation of existing models that becomes much more than the sum of its parts, similar to how narrow NLP models ‘consolidated’ into LLMs.

i.e. Multimodal++.

New work at Google DeepMind using search-enabled agents to automatically evaluate 'longform factuality' of LLMs.

'we demonstrate that SAFE achieves superhuman performance'

'We showed that SAFE is more than 20× cheaper than crowdsourced human annotators'

github.com/google-deepmin…