TLDR 2025-12-24

Welcome to Delve Shipmas. Every day this week, Delve is launching a new AI feature to make compliance more modern than ever. Yesterday, we launched the first-ever AI-native VRM platform. Today, we're launching... AI security questionnaire chrome extension.

You thought CSV's were bad. Then you opened your first platform questionnaire and saw a 200-question surprise waiting for you. Today, Delve is making that go away once and for all with our AI browser extensions for security questionnaires! Delve's agent auto-extracts all the questions, answers them based on your reports, policies, and tech stack - direclty in your Chrome tab.

See why Bland, Wisprflow, Greptile, and over 1000 more trust Delve to get and stay compliant with SOC 2, ISO 27001, GDPR and more.

Book a demo here to see the AI extension in action and get $1,500 off compliance.

Microsoft is hiring staff to translate its codebase from C and C++ to Rust. It plans to combine AI and algorithms to rewrite its biggest code bases. The engineer will be expected to work on the tools Microsoft is building to complete the project. The candidate will have to work three days a week in Microsoft's Redmond office. The role pays between $139,900 and $274,700 a year.

Many of the things we struggle to program into computers can be learned from data, but only when that data is available. If we can't learn from data, we are forced to program it in at lower performance. Computers can learn any skill, but only with enough data, and we don't want to have to need huge amounts of data for every single task. As models get better, they will need less data to learn new tasks, and when this happens, we'll finally be able to build truly general methods that combine physical understanding and cognition.

Leash Bio is a Utah-based startup that aims to build big datasets of small-molecules x protein interactions and train a model on it, then use the trained model to design a drug. The startup has a nearly pathological design to make sure its models are learning the correct thing. The startup has produced a lot of interesting artifacts from its line of research. This post looks at a few of the discoveries the startup has made.

An Australian team has discovered a way to transform normal T cells into supersoldiers in the body. The treatment helped four people with stubborn blood cancer go into remission for up to five months. The trial marks a step towards the next revolution in CAR T therapy. CAR T cells are usually made outside the body, and making them is costly for both the body and wallet. The durability of the new therapy, its longevity, and immune side effects still need to be studied, but the trials highlight the promise of a one-and-done therapy for deadly blood cancers.

Deploy freezes are a workaround, not a virtue. It's possible to evade the high-stakes peril of locks and freezes and terrifying recoveries by freezing merges and letting people work on other things, and also not freezing deploys unless your goal is to test deploy freezes. We often forget about the outages and instability we trigger with our inaction. Just run the deploy process as usual, but don't ship new code out.

The ML Systems Design interview, as an interview type, is very easy to do badly. Some companies run a parody of the Big Tech interview process, while others clearly have thought deeply about what they want, but they still don't conduct effective interviews. The interview process should be robust enough to avoid being fooled by people who are good at talking, but still accurate enough to calibrate the candidate's level as well as willingness to push out-of-band offers. The best pitch for a candidate is to present to them a slate of talented and thoughtful interviewers who could be their future co-workers, and an interview process rigorous enough to assure them that their co-workers have been thoroughly examined.

Early-career roles are traditionally more execution-heavy. Over time, contribution tends to shift toward judgment and agency. AI systems are making faster progress on execution than on defining problems, choosing what to work on, and navigating ambiguity. This disproportionately affects entry-level hiring as the execution layer becomes cheaper and thinner.

Every technological shift requires people to stop seeing the world through the past and start imagining a new one. We are still trying to build chatbots onto workflows designed for humans. What we should do instead is to imagine what knowledge work could look like when the busywork is delegated to minds that never sleep. The knowledge economy is about to undergo a significant transformation.

TLDR AI is your daily fix of LLMs, GenAI, and deep learning goodness. Same TLDR format. Still free.

Subscribe now.

If AI models were in Hogwarts houses, Gemini would be Gryffindor, Mistral would be Ravenclaw, Codex would be Slytherin, and Claud would be Hufflepuff.

Some judgment is genuinely irreplaceable, but most organizations contain far less of that judgment than their headcount suggests.

The core problems in robotics are far from solved, but most of the substantial blockers to robotics deployments now seem solvable at a technical level.

Neoen's big battery project in South Australia will be the first project in the world to deploy the new Tesla Megablock product.

The China Aerospace Science and Technology Corporation is conducting further analysis and investigation to determine why the first stage failed to be recovered.

MicroQuickJS is a JavaScript engine targeted at embedded systems that compiles and runs JavaScript programs with as low as 10 kB of RAM.