🔼 I’m not getting older, I’m just leveling up.
Late to the Party 🎉 is about insights into real-world AI without the hype.
Hello internet,
back on our usual shenanigans! This week we have lots of great content, new Python conferences and, of course, our beloved machine learning. Let’s dive in!
The Latest Fashion
- Why this data scientist found happiness in switching to data engineering
- ML-aided algorithm discovers potentially hazardous asteroid
- The Real Danger of ChatGPT by Nerdwriter1
Got this from a friend? Subscribe here!
My Current Obsession
In addition to the Daily Meeze, I talked about last week, I started trying out theming my days. Marie Poulin also has a lovely video and basic template for daily themes, and it’s been extremely useful so far. Today is my “Creative day”; this newsletter is first in line! So far, it’s working really well, but new systems always do because they’re novel and exciting. I’ll report back when I have tried it some more.
I have been working really hard on some stuff at work that we will announce at the end of the month. I am super stoked and can’t wait to share that with you.
And I’ve been playing lots of Baldur’s Gate, which is an incredible game. I don’t even like RPGs that much, and it’s just so much fun!
Most of my work these last two weeks has been personal. I overhauled my journal and some other systems I use to make my brain cooperate. The things you do as a neurodivergent person just to survive. If you have seen me within the last 5 years, you have seen someone in a deep rut with a lot of sadness and for some reason, that lifted the last weeks. I finally managed to cook for myself regularly. In fact, I haven’t ordered food in a week! I got myself some beautiful drinking glasses and the tools to eat bread for dinner (it’s the German way). I hope this isn’t just a blip, and I can keep these good vibes going. Either way, it’s nice to work, read, write, and connect again. It’s like a sunrise after a long and stormy night.
Thing I Like
I got a cute little whale bubble machine for my balcony, and except for the soapy film, it’s been so fun to sit outside with it!
Hot off the Press
I posted a silly video about getting caught in the rain, and both TikTok and Youtube loved it.
This post about the evolving role in machine learning was quite popular. My post about the frustration when you find an AI insight and it’s just smoke and mirrors was mostly ignored though.
In Case You Missed It
Don’t include these data science projects in your portfolio.
On Socials
I posted about the ML Compendium listing different ML features on Linkedin, which got a lot of traction. Mastodon quite liked my Pydata Global talk on communicating machine learning. Instagram enjoyed Stable Diffusion (go figure!).
Python Deadlines
We have new Python deadlines!
Pycon Chile CfP is closing in 27 days: https://pythondeadlin.es/conference/pycon-chile-2023/
Pydata London 2024 conference dates have been announced: https://pythondeadlin.es/conference/pydata-london-2024/
Machine Learning Insights
Last week I asked, What techniques or methodologies do you employ to measure and evaluate the effectiveness of different prompts in guiding the responses of language models? and here’s the gist of it:
Evaluating the effectiveness of different prompts in guiding the responses of language models is essential for achieving accurate and meaningful outputs. Here are some techniques and methodologies that can be employed for this purpose:
Human Evaluation:Â Human evaluators can assess the quality of model responses generated by different prompts. Evaluators provide ratings or rankings based on criteria such as relevance, clarity, and coherence of the generated text. This human judgment helps gauge how well a prompt guides the model towards desired outputs.
Quantitative Metrics:Â Quantitative metrics can measure various aspects of a generated text, such as fluency, diversity, and informativeness. Metrics like BLEU (measuring text similarity), ROUGE (evaluating summary quality), and perplexity (measuring the predictability of language) provide numerical indicators of response quality.
Comparison with Gold Standard: The model’s generated responses can be compared with a “gold standard” reference representing the ideal response. This comparison helps assess how well different prompts align with desired outputs and how close the model’s responses are to the reference text.
Adversarial Testing: Adversarial testing involves designing prompts that challenge the model’s capabilities. These prompts can include ambiguous or tricky questions, requiring the model to provide accurate and informative responses. Adversarial testing helps uncover potential weaknesses or biases in the model’s performance.
Diverse Prompts Testing: A diverse set of prompts can be used to test the model’s generalisation ability. By varying the prompts across different topics, styles, or contexts, it can be assessed how well the model responds to a wide range of inputs.
Prompt Amplification:Â Prompt amplification entails gradually refining prompts to guide the model towards desired outputs. Effective prompt formulations can be identified by iteratively modifying prompts and observing changes in response quality.
Transfer Learning Techniques:Â Transfer learning techniques can be leveraged to fine-tune pre-trained language models using domain-specific data or prompts. This allows models to be adapted to specific tasks or domains and their performance to be evaluated accordingly.
User Feedback and Iteration: User feedback on the model’s responses to different prompts is actively sought. User feedback provides valuable insights into the relevance and usefulness of generated outputs. This feedback can be used to iterate and improve the prompts over time.
By employing these techniques, how well different prompts guide the language model to produce accurate and informative text output can be assessed, enhancing the overall quality of generated outputs.
Data Stories
On average, how much time do you spend with your children?
With your partner? Alone? With Co-workers?
This visualisation takes the average US American and plots their social interactions over time.
We spend a lot of time alone and seeing how that gets higher and higher over time is quite depressing.
I guess this chart raises a big question of how we can stay connected even in old age.
Source: Rose Technology
Question of the Week
- Could you explain the concept of gradient descent and its significance in optimising machine learning models?
Post them on Mastodon and Tag me. I’d love to see what you come up with. Then I can include them in the next issue!
Job Corner
The ECMWF is hiring five people that touch machine learning right now!
Four positions in the core team to develop a data-driven machine learning weather forecasting:
- Machine Learning Engineer: Focus on model optimisation and parallel implementations to train large machine learning models on vast datasets. Prior experience with deep learning frameworks, model optimisation, and memory footprint improvements is essential. Background in earth-system modelling is welcomed.
- Observations and Data Assimilation Expert: Interface observations with machine learning algorithms and play a vital role in data assimilation. Exceptional interpersonal skills and expertise in using earth-system observations are highly valuable for this role.
- Machine Learning Scientist for Learning from Observations: Contribute to making future earth system predictions from observation data using deep learning frameworks. Experience in earth-system observation data and data assimilation approaches is desirable.
- Machine Learning Scientist for Precipitation: Specialise in accurate precipitation predictions with generative machine learning models. Experience with GANs, VAEs, or Diffusion approaches and expertise in using neural networks for precipitation prediction is advantageous.
And one on the EU project Destination Earth
You will leverage cutting-edge machine learning techniques and statistical methods to support uncertainty quantification for weather-induced extremes in the revolutionary Destination Earth (DestinE) Digital Twin. Your work will contribute to more accurate and reliable predictions, shaping the future of weather forecasting and its impact on climate understanding and resilience. If you’re a proactive and talented individual with a passion for Earth System Science and a flair for machine learning, apply now and make a meaningful difference in tackling climate challenges.
(I like to stress that these positions are, as always, written by a committee, so if identify as part of an under-represented minority, please consider applying, even if you don’t hit every single bullet point.)
Tidbits from the Web
- In case you wondered what and how rock climbing actually is
- Dimension 20 started a new season which is “Inside Out, but make it Film Noir” (including Hank Green)
- Enjoy this dog enjoying the longest puddle ever for 24 full seconds
Jesper Dramsch is the creator of PythonDeadlin.es, ML.recipes, data-science-gui.de and the Latent Space Community.
I laid out my ethics including my stance on sponsorships, in case you're interested!