Daniele Zanaga

[Daniele Zanaga @dzanaga_]

@giswqs Awesome!!! I didn't even have time to open the feature request... 😅
The speed you are building this is just incredible... 🔥🔥🔥

[Daniele Zanaga @dzanaga_]

@BasedandBraver @grok @rhaetional @misfitpatriot_ @joerogan The only possible explanation must be that Grok is Jewish, right? Very complex debate here! 🤣🤣🤣

[Daniele Zanaga @dzanaga_]

@giswqs Cool! @RubenVDKerchove

[Daniele Zanaga @dzanaga_]

@R1Quant @Burn2earn1 @MrBeast That's the point Beast is making. Tariffs are making production cheaper outside America. The opposite of what was expected. See the starting "ironically".

[Daniele Zanaga @dzanaga_]

@rocco_tanica @marcopoli17 @Lisaonthesofa Si però è trans poverina, che male fa se ogni gare vince l' oro e fa un record nuovo per la categoria femminile... Sarà un bel ricordo per i suoi figli, avere un padre che detiene tutti i record di corsa femminile mondiali over 45

[Daniele Zanaga @dzanaga_]

@dcnmforrest @GermanPolyglot Are you really so naive? He did it on purpose, not because he is a Nazi, just to troll and trigger the leftists and be under the spotlight. He's provoking the politically correct Nazis and they are all falling for it...

[Daniele Zanaga @dzanaga_]

@jeremiahscholl @IAPonomarenko Occasionally hitting the golf course with his buddy DT

[Daniele Zanaga @dzanaga_]

@PossibileIt Che qualunquismo ignorante... il patriarcato ci sarebbe in una società che considera normale e accettabile questi comportamenti. Una società che li aborra e li condanna non può essere patriarcale solo per il comportamento condannabile di un singolo. Non è un concetto difficile...

[Daniele Zanaga @dzanaga_]

@ipratese @d_boeselager Almeno se lo facevano in italia davano soldi ad aziende e lavoratori italiani...

[VITO Remote Sensing @VITO_RS_]

Greetings from Ispra 🇮🇹🤗

Proud to KO the 🆕 #Copernicus Global #Land Cover & Tropical #Forest Mapping & Monitoring service

Commissioned by the Joint Research Centre @EU_Commission the #LCFM team will provide a global service at 10m for 2020-2026 💚

👉https://t.co/7WqXMNvdio

[anton @abacaj]

Mistral just released a 87GB torrent

[Prakash @8teAPi]

Gemini Paper Flipthrough

TLDR: one step closer to AGI, but the Ultra model requires max nerfing because it’s way too dangerous to release for Google. They have complex multi-step reasoning, with tool use, search, comprehension working. Don’t know reliability, latency, cost.

> 3 models
- Ultra (GPT4 class) - efficiently scalable on TPU architecture
- Pro (GPT3.5) - cost and latency optimized version, pretraining “only took weeks”
- Nano - 1.8b and 3.25b, 4 bit quantized distilled models for low and high memory on device deployment
> Multimodal in text audio image, w video shoehorned by sampling images,
> can output text and images (I haven’t see demo of native image output, only of graph output using external tools)

Code
> Gemini powered Alphacode 2 scores 85th percentile on Codeforces vs 50% for Alphacode 1
> digging deeper AC 2 is just Gemini plus access to search and tools, ie expanding on Toolformers, and the UX work that ChatGPT has done
> it does retrieval for answers, rerank, then uses that to write code
> this is explicit leakage, as the massive search means it surely finds similar programs to solve Codeforces with rather than reasoning 😔

Graph-2-Code
> oh boy, nice functionality
> put in a graph and they can output the code that generated it
> the function recognition claim alone here is Nobel worthy if it works for complex plots (of course it doesn’t but….)

> Chart understanding -> looks like they’ve productionized the graph understanding capability first widely seen in Fuyu.
> destroys another basic human advantage for office work

Image prompting and generation
> looks just like ChatGPT-Dalle 3

Blog Generation
> it can generate a full blog post with IMAGES from a simple prompt
> omg 😱 what will all those SEO marketers do now..

IQ
> they show the basic of a visual IQ test
> too bad they didn’t eval on the ARC
> almost certainly would perform well on Raven’s matrices but that’s probably leaked into the train set

Geoguessing
> finally decided to put @georainbolt out of a job
> can’t imagine how much safety work went into nerfing this function

Homework
> this model is so good at homework it’s amazing
> don’t see the point of most homework anymore

Agentic reasoning
> it can do complex multi-step reasoning
> if reliability is good, then with search and tool use, you start to be able to do some really crazy stuff
> will be nerfed for sure

Context
> 32k context plus they actually tested retrieval accuracy within context
> seems to be claiming that Ultra model has 98% accuracy for retrieval at 32k length, which is like a 2% error rate over 100 pages of text
> also claims slow dropoff from beginning of context to end
> don’t have numbers for GPT4 but have always felt it’s closer to 70% on longer context

Evals
> the headline 90% MMLU benchmark uses Chain Of Thought with 32 samples and seems kinda dodge… but I’d allow it after all prompting should reveal mode capabilities

Other Notes
> paper includes evals up to Nov 23… this thing is hot off the research lab

[Qiusheng Wu @giswqs]

The US National Wetlands Inventory is now available as a cloud-native #geospatial format (GeoParquet) on @source_coop 🔥. A big shoutout to @jedsundwall for making it so much easy for me to contribute the first dataset to https://t.co/pmLiTpQD6V🫰Check out the link below on how to access and visualize the data using #leafmap, #duckdb, and #lonboard. Enjoy large geospatial data analysis and visualization in the cloud without have to download data to your computer 😎

https://t.co/DuIzGH1ds4

[Qiusheng Wu @giswqs]

I will be teaching #DuckDB for #geospatial starting next week. Learning spatial databases is so much easier with DuckDB. Follow the links below for updates 👇

Website: https://t.co/MiefzofnDY
GitHub: https://t.co/QbH4jTjDvM
YouTube: https://t.co/qALYaEpBqh

[VITO Remote Sensing @VITO_RS_]

🆕🛰️🚀🤩

Last night #PVCC was launched successfully with flight #VV23

With #ProbaV Cubesat Companion we aim to push affordable space technology for reliable & cost-effective #EOdata! 🌍

@aerospacelab_be made very smooth first contact after launch 🙏

➡️https://t.co/IJzNpEEEed

[Terrascope @Terrascope_BE]

🆕 Now also accessible via the #Terrascope #viewer & the #AWS Open Data Registry 👇

🌍 #WorldCover baseline 10m #landcover maps with 11 different classes

🌎 #Sentinel1 & #Sentinel annual composites

Discover more & simplify your processing 🙌➡️https://t.co/NnnWtn7sJv

[Divam Gupta @divamgupta]

Exciting news! Our team worked to make this a reality - first photorealistic podcast in VR feat. @lexfridman

This podcast was entirely hosted in VR, with realistic avatars generated through machine learning.

The immersive experience truly transports you, making it feel as if you're right there with the other person.

This is the future of communication.

[Jeremy Howard @jeremyphoward]

It looks like @johnowhitaker & I may have found something crazy: LLMs can nearly perfectly memorise from just 1-2 examples!

We're written up a post explaining what we've seen, and why we think rapid memorization fits the pattern. Summary 🧵 follows.
https://t.co/CUOWyxRJBT

[Daniele Zanaga @dzanaga_]

@rocco_tanica Gli Elohim vanno in giro con le Crocs