Paul Sinclair

As a Software Engineer - How do you feel about interviewing mainly or solely with AI? The normal and natural reaction is that many engineers are very hesitant about it. They feel firstly, like their time isn’t valued. That if a person can’t be bothered to speak with them then they’d rather not apply. And so forth. But companies using AI during the whole process is something I am actually looking forward to. And the reason for that is it removes bias and variables. You can be a talented, gifted engineer - but have you ever gone into an interview and felt that things weren’t done in a fair way? 💬 Questions were asked poorly. 🤓 The interviewer seemed more interested in showing how smart they are. 🫤 The interviewer was looking for a reason to say no, instead of a reason to hire you. This is a universal frustration. Many hiring managers are not taught how to hire. Are not vetted internally/ externally on their hiring process. It can be frankly a nightmare for everyone. Getting a job can come down to what mood the guy is in that morning, what questions they decide to ask. And so on. AI doesn’t have these issues 🤖 - if it is able to execute on the format correctly - it will remove all bias, all variables. Engineers will all have exactly the same interview. Recommendations will be made based on facts. Interviews can be reviewed and fine tuned, internal processes can be assessed/ re-assessed for fairness and accuracy. The only barrier I see is trust, and I think the way many engineers use AI these days… this is becoming a realistic option. What do you think?

13

48 Comments

I see so very many posts on here about AI changing the world, revolutionising everything, we're all going to be out of a job, you're a moron if you ignore it. Here's some brain noodles for you. AI, if you're a software engineer, is the single most well-read deep expert on every technology stack out there on your team. It's also the most inexperienced. It has little to no knowledge of your business, your legacy, why decisions were made a certain way, what trade-offs were argued over for hours in a meeting room. All it knows is that to solve problem X most people do Y. Treat it as such

21

Engineers have long praised Go for its simplicity. I used to think they were wrong. Now that's exactly why I love it. Chatting with another engineer yesterday about language choice, I realised something: the language you pick shapes how well you can work with LLM-generated code. Feature-rich languages give LLMs more rope, with more abstractions, more cleverness, and more ways to write code that compiles but hides subtle mistakes. Simpler languages flip this. Their surface area is small enough that you can actually know the whole language. When an LLM writes code, you can review it with real confidence. There's nowhere to hide. We choose languages for many reasons: performance, ergonomics, ecosystem. Now we have another one: how easily a human can reason about what the machine wrote.

54

18 Comments

Journey to 1000 models from Meta. It explains how they manage so many models in production simultaneously. The key idea is observability and automation. System Design 101. Worth reading.

118

1 Comment

I'm looking into Google's recent FunctionGemma release and it's something ... FunctionGemma is an open-model, tool-calling variant of their Gemma 3 270M which, due to its pruned size, can be run at the edge, on constrained consumer hardware, giving you all the juicy benefits of doing so (ie. privacy and 100% uptime with no vendor cost per token). I'm playing around with its 4bit quantized variant, trying to use Karpathy's approach outlined in his excellent llama2.c repo to make a bare-metal runtime for it. This is so exciting! Read more of their release here: https://lnkd.in/de-pS4vE

6

1 Comment

I think my favourite thing about LLMs is that they contain every amazing idea ever published. If you can learn how to extract these ideas you can start to actually work with them For example, today I analysed the shape of the embeddings manifold of the nomic embedding space so that I could draw a geodesic in order to interpolate smooth paths between nodes in my memory system. Could I have done this a year ago? I mean in theory but frankly I barely understand how it works now Anyway I've opened up my descartes project so if you wanna take a look (or ask Claude to) then here it is https://lnkd.in/e33TY7cn

12

7 Comments

𝗘𝘃𝗲𝗿 𝗳𝗲𝗲𝗹 𝗹𝗶𝗸𝗲 𝘆𝗼𝘂𝗿 𝗥𝗔𝗚 𝘀𝘆𝘀𝘁𝗲𝗺 𝗶𝘀 𝗴𝗶𝘃𝗶𝗻𝗴 𝘁𝗲𝗰𝗵𝗻𝗶𝗰𝗮𝗹𝗹𝘆 𝗰𝗼𝗿𝗿𝗲𝗰𝘁 𝗯𝘂𝘁 𝘂𝗻𝗵𝗲𝗹𝗽𝗳𝘂𝗹 𝗮𝗻𝘀𝘄𝗲𝗿𝘀? Often it’s not the LLM’s fault - it’s the Retriever. In a typical RAG pipeline, the retriever is like a fast librarian grabbing the top 10 chunks based on similarity. But 'looks relevant' doesn’t always mean 'really answers the query'. This is where 𝗥𝗲𝗿𝗮𝗻𝗸𝗲𝗿𝘀 come in. They are the Quality Control for your AI. --- How Rerankers Work 1. The Retriever (Stage 1): You scan 1,000 resumes for keywords like 'Python' or 'AI'. You quickly narrow it down to the top 20 candidates. This is fast and cheap but might miss a hidden gem. 2. The Reranker (Stage 2): You sit down and actually read those 20 resumes side-by-side. You look for nuance, experience, and reorder them so the #1 candidate is truly the best. Slower but more accurate. --- A Simple Example Query: 'What is the refund policy for broken items bought on sale?' Retriever finds:  • A doc about general sale items (High similarity).  • A doc about broken items (Medium similarity).  • The actual specific policy for broken sale items (Lower similarity because of word matching). Reranker fixes it: It looks at the query and the docs together (Cross-Encoding). It realizes Doc #3 is the one that actually answers the question -> moves Doc #3 to the top spot before the LLM even sees it. --- Why Rerankers Are Essential? 1. Reduce Hallucinations: If the LLM gets poor context, it will invent answers from noise. Rerankers ensure the context fed to the model is signal-rich. 2. Context Window Efficiency - Better chunks = fewer passed into the LLM, saving tokens and cost. --- The Bottom Line: If you want a production-grade RAG, you should consider - Retrieve then Rerank. ♻️ Share it with your network if you find it useful, and follow Mayank Sultania for more practical AI tips. Use top rerankers: https://lnkd.in/gRRRetDu Are you using a reranker in your pipeline yet, or are you still relying on raw vector search? Let's discuss in the comments! Image by: Simone Fuscone #RAG #AI #MachineLearning #LLM #SearchEngine #GenerativeAI #NLP #DataEngineering

19

Really insightful article on the implications of the Universal / Udio settlement. Hopefully this will lay the groundwork for future deals with musicians' consent in mind.

4

2 Comments

AI is going to take my job - or maybe not quite my job... I had an interesting task - I needed to turn a a graph (DAG) into a table so that the whole structure could be saved as a CSV. At first, easy enough, I modelled the data already and all I needed was the algorithm that does the actual job. The code I developed after a couple of hours of work just didn't look right. Having worked with Java for over 13 years helped me to develop a cetain 'taste' for what an elegant solution should look like. I wasn't happy with my code so I decided to give AI a try. Half an hour and 4 prompts later I had a perfectly working solution, that also looked elegant was concise while remaining very flexible and accomodating. I was amazed - and depressed at the same time. AI just did the job faster and better. It finally struck me, I am expendable. Most of all however, AI has just taken the whole fun of coding away from me. And I let it. Few days later we completed the sprint and I picked up the next items in the pipeline, this time... business rules - a programmer's worst nightmare. The kind of logic that is rarely concerned about the edge cases and never considers the vast number of combinations of various data points it is based on. The kind of logic that is never siloed in one specific individual or a group of people, and where every business person never really understands exactly how what they do interconnects with processes of another business domain. Yet, this is the logic that pays your wages. And it takes a human to navigate these imprecise and often contradictory requirements. This is where prompting and vibe coding contribute to the problem rather than solving it. This is where a seasoned developer makes an effort to embrace the business requirements, making decisions where data is missing or information is sloppy and imprecise. This is where a robust solution is provided back to the business for testing, with you being ready to justify all the choices made to deliver a working product. Where a potential challenge to these decisions is yet more proof you have mastered the problem, and you are well equipped to respond to these challenges in an informative and helpful way, perhaps identifying a condition the business never expected. I find it quite surprising that it is our own 'fuzzy logic' that makes us function every day, that still keeps us in charge. Though some may think we have developed AI exactly to take that 'fuzziness' from our logic away. As a developer, if you wait for the business to dictate all the requirements and answer all your questions before you write any code, I don't think you will not be in emplyment for long. Just having the brains and skills to write code is no longer enough.

12

A few takeaways I can’t stop thinking about from Boris’s perspective on building with Claude Code and what it’s doing to engineering (and everything adjacent to it): - Coding is increasingly “solved” for most day-to-day use cases. Boris hasn’t handwritten code since November—Claude Code authors it all—yet he’s still among the most productive engineers at Anthropic, shipping 10–30 PRs per day while leading a team. - The productivity gains are not incremental. Anthropic has seen ~200% engineer productivity improvement with Claude Code. The comparison that stuck with me: at Meta-scale, even heavy investment in productivity might yield a few percentage points per year—this is orders of magnitude faster. - AI is shifting from implementation to ideation. Claude isn’t just writing code; it’s starting to propose what to build by scanning feedback, bug reports, and telemetry and suggesting features worth shipping. - Engineering-adjacent roles are next. Product, design, and data science will see similar transformations as agentic AI expands into “any job where you use computer tools.” - Build for the model six months from now, not today. Design products around where capabilities are heading—even if the early product-market fit feels uncomfortable—so you’re ready to move the moment the next model lands. - Look for latent demand. Claude Code came from watching what people were already trying to do and making it easier. Cowork emerged the same way—by noticing people using these tools for non-coding tasks (e.g., MRI analysis, recovering photos from corrupted drives). - Don’t prematurely optimize token cost. During experimentation, give engineers effectively unlimited tokens. At small scale, token spend is usually trivial relative to salary. If something works and scales, then optimize. - Intentionally underfund headcount. Put fewer people on projects so teams are forced to delegate more to AI. Constraints don’t just speed things up—they change how work gets done. - The winners will be curious generalists. Cross-disciplinary builders who can frame the broader problem will outperform narrow specialists, even in an “AI-native” world. - Default to the most capable model, not the cheapest. Weaker models often cost more in the long run through mistakes and correction loops than a stronger model that gets it right early. If you’re building products or teams right now, the meta-lesson seems clear: treat AI as the new baseline for execution—and start reorganizing around discovery, judgment, and leverage. https://lnkd.in/gxNA3K7S

1

Running deep research at scale means hundreds of AI tasks concurrently searching, reading, and synthesising across millions of sources in real time. At Valyu we're obsessed with making every second of that faster. So when we noticed tasks queueing longer than they should, we dug in. The bottleneck wasn't our models or our search layer. It was how we reserved capacity across LLM providers. We went deep on this. Built a simulation environment with production data, tried four approaches, and shipped something that gave us increased concurrent capacity with zero rate limit failures. I wrote up the full technical deep dive 👇

42

4 Comments

Recently I stumbled upon a forum post from 2017 by Graydon Hoare which really peaked my interest https://lnkd.in/e-psMf_6. It was extremely insightful to see what the creator of the Rust programming language, thinks about issues beside memory safety. That modern programming languages still need to solve. Its a fascinating read for anyone interested in programming languages and programming language design or compilers/interpreters. I specifically agree a lot about how hard it is to manage errors and an error system. Its something that on my own compiler project I recently have had to start considering, as usually I like to go the if something goes wrong instantly panic route, as I am a big fan of robust and the negative programming approach. But this rarely holds up as in larger projects where in some cases you might be able to recover from errors, you need to start considering if something that you initially thought of as an irrecoverable error might for some cases become acceptable. Another example of the complexities of error management is if the path an error takes should be explicit or implicit. Like the distinction in Rust and Golang between errors (which are a return value) or something that should panic (more like classical exceptions).

3

Remember thinking "When will I ever use OS fundamentals?" in your Computer Science classes? 🤔 I just published a story that'll make you want to apologise to your Operating Systems professor. Instagram engineers discovered their Python services were mysteriously eating memory. The culprit wasn't in their application code—it was hiding in the interaction between: → Unix fork() and copy-on-write mechanics → Python's reference counting → Virtual memory management Their fix? They contributed "immortal objects" to Python 3.12—a feature that bypasses reference counting entirely and saves millions in server costs. This perfectly illustrates why OS fundamentals matter: - Performance issues often span multiple system layers - The best solutions require understanding the full stack - Sometimes you need to think like the OS to solve app problems The most interesting production bugs happen at the boundaries between your code, the runtime, and the operating system. I broke down this fascinating case study and why systems knowledge is your superpower in modern software engineering. Full story in comments 👇 What's the deepest systems-level bug you've encountered? Share in the comments! #SoftwareEngineering #SystemsThinking #OperatingSystems #SystemsDesign #Production

69

2 Comments

This week has been about Python in Node.js, modern CSS colour functions, graph queries in DuckDB and an article on LLMs explicitly ignoring ethical constraints. https://lnkd.in/eKmetcJQ

2

The AmazonQ CLI team have been smashing it recently ✨ . Have you tried Q CLI? Some of my 💖favourite 💖 features recently are: - Knowledge bases 🧠 . Some other AI tools have had this, and some rely on cli tool uses. But really powerful for making organisation knowledge available to the agent. Would love to see 'global' scoped and 'local' scoped knowledge. - Tangents 🤔 . Super useful for learning. The agent is doing something, and you want to dig and understand, but not pollute the context. Go on a tangent and deep dive! - Checkpoints 🏁 . Ever had an agent go down the wrong path? Sure you can rely on Git to change _your_ code, but it's hard to have the agent 'rollback' the context to previous states. Checkpoints let an Agent also go back in time. Combined with Todo Lists this is super powerful. Checkpoint after each todo item. - Todo lists 📒 . This feels almost table-stakes these days, but it really helps with Spec Driven Development. Came a bit later than competitors, but super useful. Use todo lists and checkpoints! - Finally... 🧑‍🏫 Because the Q CLI is open source you can see what is also being worked on. Delegated Agents are coming. Months ago Q supported custom agents, complete with their own allowed tools, system prompts and local knowledge. Having an agent delegate work to specialists will prove to be a really powerful approach. Agents for testing, coding and doing some reviews, orchestrated by another agent looking after plans and progress. Can't wait for this one. Most of these features are still behind an experiment flag. If you want to take them for a spin, make sure you turn them on with the /experiments command.

54

2 Comments

You might be paying developers to build a product nobody will ever buy. You're stuck in a trap many founders fall into. Confirmation bias creeps in when you ask friends or colleagues if your software idea is worth it. Their polite encouragement can cloud your judgment. You hear what you want to hear. Brutal honesty is thrown out of the window. At Bitsmiths Studio, I've seen countless non-technical founders make this mistake. They come to us with feature-filled roadmaps, convinced they've found a market gap. But when I ask how many strangers have pre-paid for their solution, the answer is often a disheartening zero. Building software gives you a false sense of security. You feel busy, but you're avoiding the real challenge: securing paying customers. A polished app without market demand is a recipe for failure. Time and money evaporate when your team works on a product for an invisible user base. Customer acquisition costs go sky-high when you're trying to convince people of a need they don’t recognize. It's crucial to validate your idea with cash before scaling. When partnering with founders, we focus on ensuring real market demand for their solutions. Nobody wants to build systems that cannot yield immediate ROI and encourage long-term user retention. Start validating those pain points today. Build the full product tomorrow. Remember: Validate. Build. Ship. Iterate. What steps are you taking to validate your business idea?

12

4 Comments

I was thinking about some threads on LinkedIn about the place for leetcode style interviews. I think these sort of interviews are optimised for people under 30 who are still inwardly focused and can prioritise several hours a day to drill patterns. If I think about the persona of a typical 40 year old (extrapolating from myself) they are probably spending there evenings: - spending 3 hours before work and 3 hours after work managing their kids. - filling out insurance paperwork for shortfall from their kids last stay in hospital - possibly helping an elderly frail parent with their banking, or worse dealing with their deceased estate - and for good measure maybe dealing with a random audit from the tax man for a minor claim 4.5 years ago. For this persona there really isn’t time to spend drilling leetcode, and this style of interview is probably going to filter out people who are deeply experienced having 20 years of work behind them, and have mastered juggling life and work already.

10

1 Comment