John Freeman

I've always found myself switching to Claude models when working through coding problems. There's just something about how Claude handles context and works its way through large code bases and complex logic. With Claude Code, that workflow has become ridiculously smooth.                                                    Also been experimenting with OpenAI's Codex (GPT-5 powered) and the results are way better than previous versions. The code generation quality has taken a real jump. Now here's where it gets interesting. You keep hearing stories about developers opening multiple coding agents, feeding them the same task, and manually cross-referencing outputs to find the best solution. Turns out there's an easier way. By integrating Codex as a skill inside Claude Code, you get insights from both agents in a single terminal session when you need. No tab switching, no copy-pasting, no manual comparison. Claude Code orchestrates the workflow while Codex brings its strengths. Best of both worlds in one window. Claude OpenAI

35

🚀 We’re partnering with top software engineers in New York’s early-stage startup scene. Talent is moving fast, but our clients are listening and adapting as engineers face real challenges: Interview fatigue is real. Engineers are burned out from marathon processes, long take-homes, and ghosting feedback. “We don’t get to see the product”. Companies aren’t demoing or explaining the product early on, and it’s a big turn-off. The best candidates want to believe in what they’re building. Show, don’t just tell. Tech rounds are moving up. More companies are front-loading the technical interviews to avoid time wasted later, and it’s working for both sides. Remote is shrinking. We’re seeing a definite tilt back toward onsite or hybrid. Not everyone loves it, but the shift is happening. It’s less about a coding language, more about product. Hiring managers are becoming less rigid about specific stacks, and more focused on alignment with product vision, user impact, and client collaboration. 📩 Reach out for more insights: [email protected] #Hiring #SoftwareEngineering #TechRecruiting

22

1 Comment

All this stuff about not hiring junior engineers seems so wrong-headed to me. When my non-tech friends or junior mentees ask me about AI, I liken it to VisiCalc. When VisiCalc was released (and later Lotus 123 and Excel), prognosticators predicted the end of the accounting profession. All the junior accountants who spent their days "running scenarios" by changing a cell value and manually updating all the spreadsheet values (on paper!) would lose their jobs. Then where would the next generation of senior accountants come from? What happened was, the tools got better and customers wanted to run WAY MORE SCENARIOS. Productivity went up, and all those junior accountants became super proficient in computer spreadsheets, designing them to run all new kinds of scenarios. Eventually it spawned a whole new sector we now call FinTech. Companies who stop hiring junior engineers now are just shooting themselves in the foot, and will be out-competed by those who hire junior engineers who spent their college time getting REALLY GOOD at leveraging these tools.

168

28 Comments

I wrote a short explainer about temporal AA & upscaling and the application of the Lanczos filter in 2D, which is often somewhat misunderstood. Feedback welcome! https://lnkd.in/g8Ww__uM

25

2 Comments

TOON is anti-pattern to current standards anyways. LLMs don't have large datasets on it, and you shouldn't be embedding structured data into your prompts like that anyways. 1. Run tools yourself (programmatically) 2. Insert tool results into the conversation history. These are JSON, and you can minify it! 3. Tell the model “Here is the data; now do X.” I would avoid this trend. Am I wrong?

28

22 Comments

Just submitted MAM: Maker Auction Markets to https://lnkd.in/gdfEhNhH (Designing Defi) MAM eliminates entire classes of MEV by shifting from reactive pricing to time bound collateralized commitments and user constraints/intents. Im curious do you think the way things are currently are the way things always have to be? We are what? 12-15 years into this industry? Innovation has only just begun.

1

Ran into an interesting cross-lingual hallucination with Claude recently that got me thinking about tokenization. I was prompting Claude in Mandarin about the concept of a "搭子" (a popular term for a casual activity partner, like a travel or food buddy). Instead of mirroring the term or using a valid alternative, Claude consistently responded with "搞子"—which is essentially a nonsense word. After double-checking my chat logs to confirm there were no typos in my prompts, it made me curious about what's happening under the hood. For the team at Anthropic: how exactly does Claude encode Mandarin characters? I'm wondering if this is a case of a rare token collision, or perhaps an artifact in the embedding space where the model attempts to synthesize a new compound word in a non-English language. #claude #NLP #LLM #Tokenization

6

I just published a new blog post about wordNebula, a terminal-based word processor I'm developing in C++. The inspiration came from watching George R.R. Martin discuss his use of WordStar because it simply works. Part 1 of the blog covers the Model layer, focusing on: - Choosing between Gap Buffers and Piece Tables - Designing swappable interfaces - The decision to start with the data structure before addressing the UI I wanted to create something interesting to work on. Check it out if you're into C++, architecture patterns, or curious about how text editors function behind the scenes. https://lnkd.in/gDmaUqZT

9

Agent infrastructure looks different depending on who’s building it and what problem they’re solving I stumbled upon two pieces on "agent infrastructure" one from LangChain engineers and another from Oxford/Cambridge academics. Same topic. Completely different lens. Engineering problem: “How do we actually run these things?” Agents aren’t like normal software. They run for minutes and hours, not milliseconds. They pause for human input and resume later. They need persistent state across crashes and timeouts. They need enormous amounts of compute and data. Traditional serverless and microservice architectures are not built for these workloads. Their solution: Build infrastructure with durable execution, checkpointing, task queues for bursty traffic, and real-time streaming of intermediate outputs, tool calls, traces, and human-in-the-loop systems. Academics’ problem: “What happens when agents interact with society?” When AI agents buy things online, who’s legally responsible? How do we prevent fraud? What if agents need to negotiate with each other? Their solution: Build governance protocols - identity systems, certification, rollback mechanisms. It’s like the early internet all over again. You needed TCP/UDP to make computers talk. But you also needed HTTPS/SSL/TLS to make it trustworthy. Both layers were essential. We’re doing the same thing with agents right now: - Engineers solving execution and scaling problems - Researchers solving governance problems Why this matters? Because we’re not just building better chatbots. We’re building artificial entities that will make purchases, negotiate contracts, diagnose medical conditions, handle transactions, and run autonomously in our economy. Sources: https://lnkd.in/gMjcqWyE https://lnkd.in/g4Sm6waY #AIInfrastructure #LangChain #AIGovernance #MachineLearning #AIAgents

8