Angeliki Lazaridou

Research Scientist, DeepMind

[Clément Farabet @clmt]

Next level Gemini landed - come try it @ https://t.co/Xw5ZUpe9ig

[Google DeepMind @GoogleDeepMind]

We’re releasing an updated Gemini 2.5 Pro (I/O edition) to make it even better at coding. 🚀

You can build richer web apps, games, simulations and more - all with one prompt.

In @GeminiApp, here's how it transformed images of nature into code to represent unique patterns 🌱

[Ammaar Reshi @ammaar]

Gemini 2.5 Pro (I/O edition) is amazing at coding UI ✨

I vibe coded a dictation app that:
- Structures rambling voice notes
- Removes "ums" and fillers
- Automatically formats lists

All animations and polish from one screenshot!

Demo + code below

[Angeliki Lazaridou @aggielaz]

Something tells me even the Grinch will stand up for an encore. #RAHChristmas

[Adam Liska @adliska]

I’m stoked to announce that @GlyphicAI is officially out of stealth with a $5.5M fundraise led by @p72vc to build an AI copilot for sales teams.

[Adam Liska @adliska]

We're looking for a rockstar Product Designer to join us at Glyphic. 2023 is an exciting year for us as we work with early users and iterate on the product. Apply and see more info here: https://t.co/t9EL0trAsv. (Artwork by @TomWaszkowycz)

[Angeliki Lazaridou @aggielaz]

@MarelliMar @ERC_Research Massive congratulations and looking forward to all the excitings things!!

[Angeliki Lazaridou @aggielaz]

One thing that I’m wondering is what sort of benchmarks (with easy evaluation) are out there that stress-test the same type of tasks humans pose to chatGPT. Any pointers?

[Angeliki Lazaridou @aggielaz]

@yoavgo Good point. I see semantics as a (useful) prior for modelling pragmatics. But will semantics become to pragmatics what Pos-tagging ending up being for NLP?

[Yunyao Li @yunyao_li]

👋 I will moderate the Careers in NLP panel this afternoon 5-6pm (Hall B) #EMNLP2022. We have a wonderful list of panelists.

Let me know questions and topics of most interest to you by 4pm via the following link:
https://t.co/M3FXIMudLr. I will do my best to incorporate them.

[Angeliki Lazaridou @aggielaz]

@annargrs @yoavgo The most annoying thing in papers is tuning the prose. I get a lot of research benefit from thinking about the structure of the paper and what arguments I want to provide etc. But prose is mainly convention that have little to do with research so these models can help with that?

[Kris Cao @kroscoo]

The Language team at DeepMind are hiring! Come and work with us on fundamental problems of language understanding and how this can lead us to better AI systems.

https://t.co/tQE291Zm6x

[Angeliki Lazaridou @aggielaz]

@tallinzen @_jasonwei I like that! Intuitively it seems that doing only human-task learning will require insane amount of such data. Pretraining on LM cuts this to a more reasonable amount. What’s the smallest pre-training we can do so that we can only rely on a reasonable amount of human tasks?

[Yunyao Li @yunyao_li]

❓Wondering when the #EMNLP2022 Industry Track notifications will be sent out?

We are finalizing the decisions and will send out all notification in a few hours. Stay tuned!

[Michela Paganini @WonderMicky]

Intern applications at @DeepMind for the 2023 calendar year open this Friday, September 16th! 📣

Feel free to contact me or others for info, be ready to apply, and please share with your network!

[Angeliki Lazaridou @aggielaz]

After all the great research coming up in the community around the pitfalls of static LMs, these news from @huggingface are particularly exciting! There is a great deal of applications that online LMs can enable and a lot of interesting questions around new evaluations!

[Tristan Thrush @TristanThrush]

We’re going to do it! We’ll train and release masked and causal language models (e.g. BERT & GPT-2) on new Common Crawl snapshots as they come out! We call this project Online Language Modeling (OLM). What applications or research questions can we enable or help answer? A 🧵:

[Angeliki Lazaridou @aggielaz]

@adliska Very interesting read!!

[Arthur Mensch @arthurmensch]

We're presenting RETRO at 4:15pm @icmlconf with @borgeaud_s, and later today at the poster session. Add a retrieval DB to divide your model size by 10, don't miss out!

[Angeliki Lazaridou @aggielaz]

Less than a week left for the paper submission of the inaugural EMNLP Industry Track! Paper submission June 25th. Looking forward to the exciting submissions!

[EMNLP 2026 @emnlpmeeting]

EMNLP Call for paper for the inaugural *industry track* is posted! https://t.co/PSiFmUUUYi
We have three tracks — deployed, emerging, and discovery, all focusing on real-world implementations of NLP systems. The submission deadline is July 25, 2022.

[Angeliki Lazaridou @aggielaz]

There is really a lot of great research on this very important topic across the whole community, both in terms of methods but also benchmarks. So, let’s make sure that as our models get better and bigger, they do not get stuck to a particular point in time :)

[Angeliki Lazaridou @aggielaz]

If you are at ICML check out our work on keeping QA models updated! Some reasons of why I’m super excited about this work! First, we make creative use of few-shot prompting to create a large QA dataset grounded in different points in time. Synthetic data FTW! 1/N

[Tomáš Kočiský @TomasKocisky]

Interested to chat about a new QA benchmark and how well our models adapt to new knowledge? Come see StreamingQA spotlight/poster today at #ICML2022.

By @TomasKocisky, @elenagri_, @aggielaz, @adliska, @TayfunTerzi, and others.

https://t.co/4ko3qQH0Pj
https://t.co/MOTgk7qpwG

[Angeliki Lazaridou @aggielaz]

(Vanilla) Retrieval methods are memory-intensive due to the growing search index and perform really well on low frequency items. Closed-book models otoh are compute-intensive and perform strongly on high-frequency items. This points to potential synergies between the two. :) 3/N