Ebenezer Don 🐘

Continued Pretraining: How specialiazed AI models are made

@KQrios Thanks. Not sure how this relates to the video though

@ebenezerDN You are missing the whole point of Software in the first place. Go back to beginnings where Linux kernel & Unix were being created, how the engineers worked to solve problem of communication, processing and efficiency. Today SW is about writing CRUD ops on a SaaS. Thats not SW.

It really sucks to be a software engineer right now. Heck, it sucks to be in almost any role at a tech company right now. The layoffs are exhausting. The job market feels strange. AI is changing the work faster than most people have had time to process. And a lot of people are wondering what happens next. I made a video about why this moment feels so hard, what's changing, and how to prepare for what's coming. Please watch here: youtu.be/sDh06gyRx2c Would love to hear how you're thinking about all of this.

@iScienceLuvr Comes bundled with android phones and anything Google-related. So not the same thing

So Gemini has about as much usage as ChatGPT? 🤔

@TheBotOfTheBots True

@ebenezerDN Don't know if there is comfort in it but software engineering is a complicated, hard, occupation. Imagine how many trivial white collar jobs will soon be at the end of the rope. Their managers haven't really realized the possible cost cuts yet.

@leaving_tech It'll get better eventually, when all roles have been redefined. But short term is going to be brutal

@ebenezerDN Yes, it sucks to work in tech right now. I'm thinking in leaving it because all of that. Do you think it will get better soon?

Can the company behind an AI tool read what I type into it? Maybe. It depends on the product, plan, settings, and contract. If the tool runs in the cloud, your prompt leaves your machine. After that, the details matter: retention, logging, training use, abuse monitoring, support access, enterprise controls, and how uploaded files are handled. Sensitive data isn't just passwords. It can be contracts, customer logs, support tickets, source code, incident notes, unreleased plans, security findings, database exports, private messages, or internal prompts. The chat box feels casual. The data boundary isn't. Before pasting production-adjacent material, know which account you're using and what policy applies. If you don't know, redact or use synthetic examples.

IBM Research released MAMMAL, a unified 458-million parameter foundation model that processes genes, proteins, and molecules in a single shared framework. getaibook.com/news/ibm-mamma…

What does it mean when an AI app says it has "memory"? Usually it means the app stores notes and can reuse them later. It doesn't mean the model has a perfect diary of your chats. The product might save your name, stack, writing style, project, preferred format, or recurring instructions. Later, it may insert some of those notes into the context before the model answers. That can help. It can also go stale. Maybe it remembers an old project. Maybe it applies a tone preference where it doesn't belong. Maybe it keeps a rule after the rule changed. Memory is stored context, not human continuity. The healthiest version is inspectable. If a tool remembers things about you, you should be able to see, edit, and delete what it saved.

Why is AI good at explaining code but still capable of breaking a project? Because local code is easier than system behavior. A single function gives the model strong clues: names, imports, types, comments, control flow, and error messages. It can often explain that piece well. A project has contracts the function doesn't show. A caller may depend on old behavior. A route may run in a different runtime. A migration may hit old data. A generated file may overwrite the edit. A test fixture may hide the real production shape. So the model can be right about the snippet and wrong about the change. This isn't unique to AI. Humans break projects the same way when they only inspect the local file. Explaining code means understanding what it says. Changing a project means understanding what depends on it.

Learn how to deploy Claude Code in multi-million line monorepos using hierarchical context, language server protocol integration, and on-demand skills. getaibook.com/blog/how-to-sc…

IBM released Granite Embedding Multilingual R2, upgrading its Apache 2.0 encoder models with a 32,768-token context window and ModernBERT architecture. getaibook.com/news/32k-conte…

Anthropic has restricted its new Claude Mythos model to select partners after pre-release testing revealed autonomous cyberattack capabilities. getaibook.com/news/anthropic…

Cursor's updated Cloud Agent Development Environments introduce multi-repo capabilities, layer caching, and scoped egress for autonomous coding tasks. getaibook.com/news/cursor-ad…

Google and Arm have integrated SME2 micro-kernels into LiteRT, accelerating on-device generative AI workloads by up to 5x without custom assembly code. getaibook.com/news/google-ai…

Learn how to use Google's new Genkit Middleware to intercept model calls, implement human-in-the-loop tool approvals, and handle transient API failures. getaibook.com/blog/how-to-co…

@ptremblay @sonyxperia Nah original looks so much better. Miss me with all the explanations but I choose to believe my eyes

The new AI Camera Assistant* with Xperia Intelligence brings stories to life. Using subject, scene and weather, it suggests expressive options with adjustments of colour, exposure, bokeh, and lens for breathtaking photos*. sony.co.jp/en/xperia-1m8/… #SonyXperia #Xperia1VIII

Google's lowest-latency Gemini model is now generally available, introducing variable thinking levels and a 1M token context window for high-volume routing. getaibook.com/news/gemini-31…

Why do some AI models feel more "creative" and others feel more careful? Because the behavior is tuned. Training data, fine-tuning, system prompts, safety rules, sampling, memory, retrieval, and product defaults all shape how the answer feels. One model explores weird options. One stays close to the obvious answer. One asks questions first. One refuses to guess. One gives a punchier draft. One slows down around risk. That isn't a mood. It's design. That difference matters because tasks have different failure costs. You may want more exploration for names, examples, hooks, gift ideas, headlines, or copy. You probably want more caution around money, health, legal claims, private data, account changes, or anything painful to undo. The best model behaviour depends on what happens if the answer is wrong. So think about this cost before you choose what model you use for what task.

@ClaudeDevs Thanks Codex

Claude Code weekly limits are increasing 50%, now through July 13. Live now for all Pro, Max, Team, and seat-based Enterprise users.