[September 2025] AI & Machine Learning Monthly Newsletter 💻🤖

69th challenge! In the event you missed them, you may read the previous issues of my monthly A.I. & Machine Learning newsletter here.

Hey there, Daniel right here.

I’m an A.I. & Machine Studying Engineer who additionally teaches the next beginner-friendly machine studying programs:

I additionally write usually about machine studying by myself weblog in addition to make movies on the subject on YouTube.

Since there’s lots happening, the utmost care has been taken to maintain issues to the purpose.

Sufficient about me! You are right here for this month’s A.I. & Machine Learning Monthly Newsletter.

Sometimes a 500ish (+/-1,000ish, normally +) phrase put up detailing a few of the most attention-grabbing issues on machine studying I’ve discovered within the final month.

Here is what you may need missed in September 2025 as an A.I. & Machine Studying Engineer… let’s get you caught up!

My work

• Birthday article: I turned 32 firstly of September and wrote an article on some things I’ve realized (I do the same article yearly).
• Work in progress: I’m engaged on an LLM fine-tuning tutorial. Inside, we’ll take a small LLM and fine-tune it to carry out a particular activity. I’ve bought the define prepared, now I simply have to put it collectively in a enjoyable method.
• Coming very quickly: My Hugging Face Object Detection Project Course is veryyyy near being launched on ZTM. A quote from the editor: “classes needs to be completed in subsequent couple days!”

From the Web

Post Training 101 by Han Fang and Karthik Sankararaman

Ever surprise how an LLM reads the web however then gives you helpful responses?

Pre-training offers with the artwork of subsequent token prediction (e.g. “The fox jumped over the ___”).

However Submit-training takes the realized illustration of token ordering and helps steer it in the direction of one thing people need/is helpful.

The Submit-training 101 information walks you thru examples of how this occurs, together with:

• From next-token prediction to instruction following
• Supervised Wonderful-Tuning (SFT) fundamentals
• Reinforcement studying (RL) methods similar to RLHF, RLAIF, RLVR
• Analysis methods for assessing mannequin high quality

I liked the stream of this weblog put up.

It’s a superb information with loads of knowledge examples for every approach.

Instance of a Supervised Wonderful-Tuning (SFT) instance for put up coaching a LLM. SFT includes exhibiting the mannequin examples of inputs and direct desired outputs. Supply: Submit-training 101 weblog put up.

Certainly one of my favorite quotes from the put up was a reference to the Gemini 2.5 Professional paper:

The Gemini 2.5 Professional paper particularly emphasised that “Because the preliminary announcement of Gemini 1.5, important developments have been made in our post-training methodologies, pushed by a constant deal with knowledge high quality throughout the Supervised Wonderful-Tuning (SFT), Reward Modeling (RM), and Reinforcement Studying (RL) phases.

After studying this information, you’ll have an important reference level to lots of the methods which go into making a world-class mannequin.

Hugging Face TRL now supports native vision fine-tuning

If you want to fine-tune an LLM (or SML – Small Language Mannequin) or VLM, Hugging Face’s TRL (Transformers Reinforcement Studying) is one among your finest choices.

And with a latest replace, TRL now natively helps imaginative and prescient in addition to language fine-tuning.

Help consists of the next RL (Reinforcement Studying) strategies:

• Blended Desire Optimization (MPO)
• Group Relative Coverage Optimization (GRPO)
• Group Sequence Coverage Optimization (GSPO)
• Direct Desire Optimization (DPO)

TRL additionally helps Supervised Wonderful-Tuning (SFT).

Right here’s a brief instance of fine-tuning a VLM with TRL:

from trl import SFTConfig, SFTTrainer
from datasets import load_dataset

coach = SFTTrainer(
    mannequin="Qwen/Qwen2.5-VL-3B-Instruct",
    args=SFTConfig(max_length=None), 
    train_dataset=load_dataset("trl-lib/llava-instruct-mix", break up="prepare"),
)
coach.prepare()

Now when you have your individual knowledge, you may simply customise and open-source mannequin to be much more aligned with your individual distinctive downside house.

You possibly can then deploy this mannequin at your individual discretion with out having to fret about third-party APIs.

See the instance sft_vlm.py script for extra.

Bonus: Talking of fine-tuning, right here’s a Tweet from Nathan Lambert, writer of The RLFH E-book and researcher at Ai2 with reference to utilizing RL-based fine-tuning versus SFT. In my expertise, small fashions fine-tuned on particular knowledge can equal or exceed the efficiency of these a lot a lot bigger (which haven’t been fine-tuned).

Rule of thumb: If in case you have a mannequin below 15B parameters, strive supervised fine-tuning. In the event you’ve bought an enormous mannequin, go RL. After all, precise expertise many range. Greatest to experiment, experiment, experiment! Supply: Nathan Lambert on X.

Ethan Ding writes about agents and clouds

Do you construct an AI agent after which a cloud?

Or do you construct a cloud after which an agent?

OpenAI has an agent however no cloud (but).

Google has each.

A number of startups are in between like Bolt and Loveable as customers of brokers.

And Netlify and Supabase as suppliers of cloud companies.

Anthropic is making an attempt to go vertical with their Claude mannequin in addition to all of the instruments to go alongside it.

As highly effective as fashions get, you continue to want a spot to retailer knowledge (a database) and a spot run to utility code (internet companies).

Brokers (yellow) vs Clouds (inexperienced) and all the things in between. Supply: Ethan Ding weblog.

Modal releases notebooks — GPU-powered notebooks in seconds

Talking of brokers and clouds, Modal is a quick compute supplier from 1 GPU to many in seconds.

They usually simply launched assist for Jupyter Notebooks hosted on Modal.

Meaning you can begin up a Jupyter Pocket book occasion with a number of GPUs on the backend in a couple of seconds and have it robotically energy down while you’re not utilizing it (the default is half-hour of idle = shutdown).

I used to be in a position to go from zero to operating Qwen3-4B on a H100 GPU in ~2 minutes.

That is actually useful to experiment with a smaller GPU regionally and even on Google Colab after which simply improve the {hardware} utilizing Modal when mandatory.

Modal payments per second so that you solely get charged when the pocket book is operating.

Instance interface of utilizing Modal notebooks with a H100 on the backend. Supply: Modal documentation + my very own use case.

Case study: Hugging Face Is All You Need

A cool learn from the Finegrain group (they make high-quality picture enhancing instruments) on how generally easy is finest in terms of bettering mannequin high quality.

They share how they use the Hugging Face ecosystem to create:

• A easy internet app (utilizing Hugging Face Areas) that lets human testers play with the Eraser mannequin — decide a picture, brush over an object to erase, and thoroughly examine the consequence.
• A approach to report any points (utilizing Hugging Face Datasets) to file the inputs/outputs and describe what went improper from a top quality perspective.

Doing this implies they will take a look at a brand new mannequin and use the found points to enhance it over time.

It is a comparable workflow to what I’ve not too long ago utilizing for an object detection mission with a consumer: prepare a mannequin, add to Hugging Face Areas, strive it out, enhance it with higher samples within the subsequent coaching run, add, take a look at, repeat.

An experiment loop is all you want.

Case study: TensorZero finds out you can save 30x on costs and 4x on inference time by fine-tuning models

In the event you’ve bought a particular use case with loads of knowledge, chances are high, fine-tuning a smaller mannequin will prevent on money and time.

One other workflow is to make use of the perfect mannequin you may by way of API, then fine-tune a smaller mannequin (or the very same mannequin, if it permits) to repeat the identical activity.

They examined Gemini 2.0 Flash, Gemini 2.0 Flash Lite, Qwen3-8B, GPT-4.1 nano and GPT-4.1 mini (each zero-shot and fine-tuned variations).

In lots of zero-shot settings, GPT-4.1 was the perfect performing.

Nonetheless, after fine-tuning, nearly the entire fashions carried out on par or higher than GPT-4.1 on a number of benchmarks leading to price financial savings and quicker inference time.

Each mannequin examined achieved important price financial savings after fine-tuning in comparison with the unique GPT-4.1 mannequin. Supply: TensorZero weblog.

Thinking Machines discovers and publishes how to make LLMs deterministic

If you wish to be taught extra about GPU programming, learn all the things Horace He has printed.

See earlier works similar to gpt-fast and *Making Deep Learning Go Brrrr From First Principles.*

And his latest put up, Defeating Nondeterminism in LLM Inference, in collaboration with Considering Machines is sensational.

In the event you’ve used an LLM, you would possibly’ve discovered that given the actual identical enter, it produces a special output.

Even when operating regionally and setting temperature to 0, this may nonetheless occur.

This may be useful generally however while you’re making an attempt to run repeatable assessments, determinism is your pal.

It seems, this isn’t a trivial downside to unravel.

Till…

From the write up:

In different phrases, the first purpose almost all LLM inference endpoints are nondeterministic is that the load (and thus batch-size) nondeterministically varies! This nondeterminism will not be distinctive to GPUs — LLM inference endpoints served from CPUs or TPUs may also have this supply of nondeterminism.

Whenever you ping an LLM endpoint, chances are high your inputs are within the queue with another person.

So whereas your requests is perhaps deterministic, you may’t assure the entire others (together with yours) are.

Supply: Considering Machines weblog.

How do you repair it?

You match three operations batch invariant (it doesn’t matter what’s within the batch): RMSNorm, matrix multiplication and a focus.

See the remainder of the weblog put up for this works in follow in addition to example code on GitHub for how to make it work.

Stephan Tulkens blog is an excellent read for more on tokenizers

Tokenizers flip sequences of information similar to textual content, audio or pictures into numbers so machine studying fashions can discover patterns in them.

They’re step one and second final step in most interactions with LLMs.

For instance, “Good day my title is Daniel” would possibly get was [hello, my name, Dan, iel] (I made this instance up) after which would get mapped to numbers similar to [43, 5556, 339, 46, 119] (once more, made up).

It seems that some tokenizers are cased (they keep in mind capitals, “Dan” is completely different to “dan”) and a few are uncared (“Dan” is identical as “dan”).

Relying the way you enter textual content into your tokenizer, this may affect efficiency.

The excellent news is that you could Stephen exhibits methods to mitigate this.

Anthropic release a guide on prompt engineering vs context engineering

In final month’s AI & Machine Learning Monthly (August 2025), we lined immediate engineering vs. context engineering.

The brief model is:

• Immediate engineering = normally 1 step, for instance, immediate in, textual content out
• Context engineering = will be a number of steps, earlier steps affect future steps, add in paperwork, instruments and extra

Anthropic’s information nails down a few of these phrases.

I particularly just like the query it asks proper firstly:

What configuration of context is more than likely to generate our mannequin’s desired behaviour?

After which narrowing down on what good context engineering is (daring is mine):

Good context engineering means discovering the smallest attainable set of high-signal tokens that maximise the chance of some desired consequence. Implementing this follow is far simpler mentioned than completed.

Some sensible suggestions:

Offering examples is a widely known finest follow and one thing we proceed to advise.

And at last, a easy definition for brokers:

LLMs autonomously utilizing instruments in a loop.

Immediate engineering normally includes a easy enter and output workflow. Whereas context engineering will be considered a system which mixes 1 to N steps and instruments. Supply: Anthropic weblog.

Daniel’s Open Supply AI of the month

Qwen causes an avalanche

The Qwen group are effectively in entrance for the unofficial “Open Supply AI group of the yr” trophy 🏆.

It appears nearly each month they’re releasing frontier fashions with open-source licenses throughout nearly each area.

This month is not any completely different:

• Qwen3-Omni is a sequence of fashions which may function in all modalities: audio, video, pictures and textual content. At present achieves the perfect outcomes on 32/36 speech recognition duties.
• Qwen3-Next is a sequence of fashions which balances efficiency with variety of parameters. Utilizing 80B parameters with 3B energetic, the fashions are in a position to carry out on par or higher than Qwen3-235B-A22B, a mannequin with ~4x extra parameters and ~6x extra energetic parameters. It additionally sees an unimaginable inference pace efficiency in comparison with Qwen3-32B-Base attributable to having ~10x much less energetic parameters. I can verify… In my few trials utilizing the the Hugging Face Inference API… they’re quick!
• Qwen3Guard is a sequence of fashions which assist defend from unsafe prompts. For instance, you would put the Qwen3Guard mannequin in between your customers enter and your manufacturing system.
• Qwen3-VL is a sequence of fashions which marries collectively Qwen’s flagship text-based LLM in addition to a imaginative and prescient encoder for unimaginable efficiency within the imaginative and prescient modality. The mannequin is on par with or higher than Gemini 2.5 Professional’s and GPT-5’s imaginative and prescient capabilities. It my transient hands-on expertise, I’ve discovered it implausible at localization and OCR (see picture beneath). Learn the Qwen3-VL blog post for extra or see the Qwen3-VL cookbooks series on GitHub for concepts.

Hooooweeee! If they will preserve this momentum up, it’s going to be an enormous end to the yr.

Instance of utilizing Qwen3-VL for structured knowledge extraction. The mannequin is able to simultaneous localization (detecting bins, on the left), in addition to detecting textual content values. I then used the identical mannequin to show the acknowledged textual content into markdown (picture on the fitting). Seems prefer it bought all of it right! Supply: Creator creator.

Google DeepMind Release EmbeddingGemma, a 308M parameter embedding model

Embedding fashions are designed to show textual content into realized representations which will be in contrast and used for duties similar to retrieval.

For instance “largest international locations on the planet” would possibly get was [0.234, 0.934, 0.004…].

EmbeddingGemma does this with sensational efficiency at a really manageable dimension.

Some fast stats:

• 100 languages
• Highest rating embedding mannequin below 500M parameters
• Sufficiently small to run on gadget
• 2k context window (embed sequences as much as 2k tokens lengthy)
• Can leverage Matryoshka Illustration Studying (MRL) to transform embeddings to dimension 768, 512, 256 or 128
• Quantization-Conscious Coaching (QAT) utilized throughout fine-tuning so the fashions will be quantized (made even smaller) with out massive efficiency loss

Word: You have to be conscious of the immediate directions. Every activity has a particular immediate instruction, for instance, to embed a question of “largest international locations on the planet”, it’s best to prefix “activity: search consequence | question: {content material}”, for instance, “activity: search consequence | question: largest international locations on the planet”. See the EmbeddingGemma model page for extra.

If you want to construct a RAG (Retrieval Augmented Technology), it’s best to take a look at EmbeddingGemma and see the way it performs.

See the Google Developer breakdown or read the research paper for extra.

Apple release MobileCLIP2

MobileCLIP2 is a mannequin designed to match pictures and textual content.

CLIP stands for “Constrastive Language-Picture Pretraining” so pictures and texts get encoded into the identical embedding house.

The MobileCLIP2 fashions have been designed to run on cellular units similar to iPhones (therefore “cellular” of their title).

This implies they’re light-weight and run fassssssssst, they’re able to operating from between 3ms and 30ms on an iPhone 12 Professional.

It’s extremely seemingly that the MobileCLIP2 fashions are what energy the search characteristic on Apple’s Pictures app, permitting you to seek for issues like “Georgia standing in entrance a e book shelf”.

Get the MobileCLIP2 models on Hugging Face, check out a MobileCLIP2 demo on your own images and skim the research paper for extra.

Since CLIP-like fashions are skilled to match texts and pictures, the higher your textual content matches a picture (and vice versa), the upper rating it is going to get. And since MobileCLIP2 has seen many various sorts of pictures and texts, it will probably even match its personal outcomes graph fairly effectively. Discover the extra descriptive texts get the next rating. Supply: MobileCLIP2 demo.

5TB of high-quality vision data, ready for training via FineVision

Probably the most underrated abilities on the planet of AI is with the ability to curate a dataset.

Two issues appear to be true: scale is essential (extra samples are usually higher) and task-specific samples are essential.

That’s mirrored within the new FineVision dataset from Hugging Face.

Some fast particulars:

• 24 million samples
• 200 datasets mixed into the same interface
• 17M pictures
• 89M question-answer turns
• 10B reply tokens
• 5TB of high-quality knowledge

The researchers quote their efforts as:

FineVision was a large act of information curation. We began by accumulating publicly out there datasets, and augmenting underrepresented classes. We then evaluated all datasets for duplicated knowledge internally and benchmark contamination. This knowledge is then cleaned and rated, earlier than being added to the ultimate combination.

It seems it was value it, a nanoVLM mannequin skilled on the FineVision dataset outperforms different nanoVLM fashions skilled on different comparable however smaller datasets:

Rankings of various nanoVLM situations skilled on varied picture + textual content datasets. FineVision-trained fashions begin out slower however find yourself being the perfect ranked general with longer coaching. Supply: FineVision weblog put up.

Lastly, one other attention-grabbing perception based mostly on whether or not it’s essential to solely prepare on “top quality” samples (samples with a median ranking of X or above) or to easily prepare on all the things:

Merely coaching on essentially the most various knowledge, that one containing all samples, outperforms in benchmarks (Fig. 6) (Fig. 7). This might imply a number of issues. Firstly, we are able to see nearly the identical distribution within the ranks throughout all filters: from finest to worst with a rise within the ranking threshold. For instance the visible dependency and the picture correspondence ranking each lead to precisely the identical distribution of rankings, comparable to the pure order of choices, 1 via 5. This might point out that with a sufficiently massive dataset that you just prepare on lengthy sufficient, it hurts extra to take away samples, even when they had been judged to be of low high quality, than to coach on them.*

It appears the mix of each scale and top quality samples is what leads to the perfect performing mannequin.

Ettin encoder and decoder pairs of models outperform ModernBERT

Lately LLMs have favoured decoder fashions.

However encoder fashions are nonetheless extremely helpful.

When constructing the Ettin sequence of fashions, the researchers discovered:

The outcomes present clear patterns:

Encoders dominate classification and retrieval: On MNLI classification, even a 150M encoder (89.2) outperforms a 400M decoder (88.2). For retrieval duties, the hole is smaller however nonetheless noticeable – particularly when decoders should not skilled with MNTP.

Decoders excel at era: On generative duties, decoders keep constant benefits, with the efficiency hole truly widening at bigger mannequin sizes.

Dimension would not at all times matter: A 400M encoder beats a 1B decoder on classification duties, whereas a 400M decoder beats a 1B encoder on era duties.

The excellent news is, for each encoder within the Ettin sequence, there’s an identical decoder.

The one distinction is the coaching goal.

Encoder fashions use bidirectional consideration, permitting every token to “see” all different tokens within the sequence.

Decoder fashions use causal consideration, the place solely earlier tokens are seen to allow autoregressive era.

The Ettin fashions come within the following sizes: 17M parameters, 32M, 68M, 150M, 400M and 1B.

An ideal assortment of sizes to run on smaller units with out compromising on efficiency as every performs on par or higher than an equal mannequin of its dimension.

See the Hugging Face collection, research paper and GitHub for more.

TinyLettuce Encoders show how powerful small specific models can be

Starting from 17M-68M parameters (that’s million fairly than billion), the TinyLettuce fashions punch effectively above their weight, acting on par or higher than a lot bigger fashions similar to gpt-oss-120b and Qwen3-235B on a hallucination detection activity.

Given a context, query and reply, the fashions predict at a token stage which tokens within the reply is perhaps hallucinated.

They obtain these outcomes by creating task-specific artificial knowledge after which fine-tuning a set of base Ettin encoders (see above) on it.

The fashions are even sufficiently small to run on CPU.

In the event you’ve bought a big scale repeatable particular activity, I’d extremely suggest trying out the TinyLettuce weblog put up for concepts you may reproduce in your individual workflow.

Christmas for OCR fashions!

• MinerU2.5 is a 1.2B VLM mannequin targeted on OCR able to parsing tables and mathematic formulation. It may function at 2.12 fps on an A100, AGPL-3.0 licensed.
• POINTS-Reader is a fine-tuned model of Qwen2.5-VL-3B able to outputting markdown and HTML from paperwork. The paper exhibits a superb methodology for bootstrapping paperwork for pretraining. Apache 2.0 license.
• PP-OCRv5 and PP-StructureV3 are the newest variations of the PaddlePaddle doc recognition fashions. These fashions are far smaller than most (within the vary of lower than 100M parameters). The library is sort of fleshed out on GitHub. Accessible below Apache 2.0 license.
• Granite-Docling-258M is a mix of SigLIP2-Base-512 in addition to an LLM (Granite 156M). The mannequin is designed to soak up paperwork and out the docling format which gives gadgets in particular tags, similar to , , in addition to their places in tags. Educated utilizing the nanoVLM framework and out there below Apache 2.0, strive the demo on your own documents.