ML for SWEs 1: Apple Pushing Their AI Back Isn't as Bad as You Think

Machine learning for software engineers 3-7-25

The hardest part about working in AI and machine learning is keeping up with it all. Each week I package everything software engineers should know about what happened in AI into a digestible package with an emphasis on why those things are important. Subscribe to join 5000+ engineers receiving these emails directly in their inbox.

---

Side note: We’ve got a new logo! Let me know what you think of it.

Here are the AI developments from this past week you should be aware of. The most important development to understand is under the GIF.

GPT-4.5 was released to mixed reactions. It isn’t considered a frontier model and was met with a lot of disappointment. GPT-4.5 is still an important release because OpenAI’s focus wasn’t on pushing performance but instead on developing model character. The release has again raised concerns about the limit of scaling performance via pretraining. More resources on this below.

Mistral released Mistral OCR, an Optical Character Recognition API. This API combines state of the art models, RAG, and other software elements to achieve impressive performance on multi-modal document understanding. This has potential to improve many software pipelines where text extraction is important.

We got to see Claude play Pokemon live. It was interesting to watch where the AI struggled and to understand more about how it was done. Claude code was also made available to everyone. These two things come as Anthropic raised a Series E of funding coming in at a valuation of $61.5 billion. AI development is expensive so a lot of money is needed, but that’s a massive valuation.

Most important for software engineers to understand is Apple delaying it’s release of a revamped, completely AI-based Siri. Apple is clearly behind in this space, but to anyone who understands the AI product landscape this shouldn’t come as a surprise.

Apple isn’t an AI company. Their direct competitor for a conversational AI assistant is. Thus, Apple is compared to a company that has had machine learning infrastructure put together for over a decade.

Putting together machine learning infrastructure is incredibly resource intense. It requires a lot of time, money, and engineering talent to get right and it’s even more difficult when the AI is being used in a consumer product. Both the model and the infra need a level of polish and consideration that isn’t necessary for research. Since Apple hasn’t previously been an AI company, they’re having to get all of that going at once right now.

To put this into perspective, both OpenAI and Google took years to bring their AI to products. Google Brain was founded in 2011 and we didn’t the first real consumer product using hardcore AI until 2016. OpenAI was founded in 2015 and their first real consumer product was ChatGPT, released in 2022. Between founding and their first product, both of these companies released research previews and built models, just as Apple is currently.

Along with the other considerations, Apple is a company that likes to get things very polished before releasing them. This is an important part of their brand image—things “just work”. Apple Intelligence wasn’t met with the same enthusiasm as their other product releases and I’m certain executives don’t want a repeat of that reception with another AI product.

In my opinion, it’s good on Apple for making this decision. While not an ideal situation to be in, the fallout of pushing AI product releases back is much less than the potential fallout of botching the release altogether. Apple also doesn’t need AI to keep the business afloat.

As a software engineer, it’s important to understand business decisions because they drive the changes we make. Understanding the nuance of the decisions greatly improves the impact you can have on the business you work for (or in). Many people judge only headlines and took this event as Apple’s AI being dead. Society’s Backend readers won’t be those people.

Below are the important resources for the week. If you want to learn ML fundamentals for free, check out my roadmap. If you missed last week’s updates, check it out below. If you want to support Society’s Backend and get access to even more AI and ML resources, you can subscribe for just $3/mo.

The most important resources this week

On the US AI Safety Institute by

Dean W. Ball

: The US AI Safety Institute (AISI) should be preserved and expanded to effectively assess and mitigate the risks posed by advanced AI systems, especially as their capabilities evolve and the potential for catastrophic harm increases.

AI by Hand Workbook (Print on Demand) – AI BY HAND by

Tom Yeh

: The AI by Hand Workbook offers 300 original exercises by Prof. Tom Yeh to help build a strong mathematical foundation for deep learning concepts.

Why You Should Never Let AI Debug for You by

Logan Thorneloe

: Never let AI debug for you, as the tedious tasks in software engineering are crucial for skill development and understanding, while AI should instead be used as a collaborative tool in the coding process.

The State of Machine Learning Competitions: In 2024, machine learning competitions saw a resurgence with platforms hosting diverse challenges and significant prize pools, including notable events like the AI Mathematical Olympiad and ARC Prize, reflecting advancements in AI reasoning and participation from various organizations.

Expanding AI Overviews and introducing AI Mode: Google has launched AI Mode and updated AI Overviews to enhance search capabilities, making it easier for users to get high-quality responses to complex queries.

AI Policy Primer (February 2025) by

AI Policy Perspectives

: Recent developments in AI policy include significant investments from France and the EU in AI infrastructure, a proposed market-based approach to enhance AI safety, and a gradual increase in the deployment of autonomous vehicles despite ongoing challenges.

Cloud Infrastructure for AI: What You Actually Need to Know [Guest] by

Devansh

: Understanding the critical components of cloud infrastructure—including compute, storage, and networking—is essential for leveraging AI effectively and optimizing performance in various applications.

The Insurmountable Problem of Formal Reasoning in LLMs by

Alejandro Piad Morffis

: Large language models (LLMs) fundamentally struggle with formal reasoning due to intrinsic limitations in their architecture and probabilistic nature, which prevent them from achieving reliable, provably correct logical deductions.

GPT-4.5 Feels Like a Letdown But It’s OpenAI’s Biggest Bet Yet by

Alberto Romero

: GPT-4.5 is OpenAI's largest AI model to date, but it is seen as a disappointing step backward in performance that may ultimately serve as a foundation for future advancements.

Mercury LLM Matching Claude and GPT is more Important than You Think [Breakdowns] by

Devansh

: Mercury's Diffusion Large Language Models (dLLMs) demonstrate exceptional speed and quality in text generation, outperforming traditional autoregressive models like GPT and Claude while offering greater control and versatility across various data types.