Hacker News: jsemrau

New comment by jsemrau in "Ask HN: What Are You Working On? (July 2026)"

jsemrau — Mon, 13 Jul 2026 02:05:11 +0000

Code World Models in Simultaneous Move settings like Capital Markets. DeepMind's CWM approach relies on standard MCTS/IS-MCTS, which assumes a single active player at each node.

This doesn't work in simultaneous-move settings like Orbit Wars (or order-book markets), converging to an exploitable pure strategy rather than a Nash equilibrium.

LeCun's JEPA, by contrast, is a learned neural world model, which lacks the determinism, speed, and debuggability of a code-based simulator. Thus, it can drift or predict illegal states, and you can't inspect why it made a prediction the way you can trace a Python function.

TL;DR: The benefit is better auditability and easier RL-like training. The SM-MCTS extension fixes the first problem (decoupled UCB per player approximates Nash equilibrium instead of a pure strategy) while keeping the second advantage intact (a deterministic, inspectable code simulator).

https://github.com/ternary-ai/ow-code-world-model https://jdsemrau.substack.com/p/a-self-improving-code-world-...

New comment by jsemrau in "Lost city discovered beneath Egypt's desert with ancient church"

jsemrau — Sat, 11 Jul 2026 11:32:41 +0000

Like the rest of the world. Since there are many non-Western countries that still have slaves there is no original sin.

New comment by jsemrau in "Lost city discovered beneath Egypt's desert with ancient church"

jsemrau — Sat, 11 Jul 2026 11:31:31 +0000

I don't believe in the original sin of slavery in Western countries. Slavery has been existed for centuries before and, despite being abolished by Western countries (hint: there were several wars were people lost lives), still exists in China, India, Indonesia, and a large group of Arab countries.

New comment by jsemrau in "Lost city discovered beneath Egypt's desert with ancient church"

jsemrau — Sat, 11 Jul 2026 07:25:36 +0000

There are more people enslaved now than at any point in time in history. Guess the countries.

New comment by jsemrau in "Lost city discovered beneath Egypt's desert with ancient church"

jsemrau — Sat, 11 Jul 2026 05:50:51 +0000

Please educate yourself more on this. Step out of your bubble. Christian nations have created the cultural and technical achievements over the last thousands of years. There is nothing even remotely comparable.

New comment by jsemrau in "Local, CPU-Friendly, High-Quality TTS (Text-to-Speech) with Kokoro"

jsemrau — Tue, 07 Jul 2026 23:13:38 +0000

I am entirely sold on Qwen3-TTS's voice cloning. It runs locally and I can run it as many times as I want.

New comment by jsemrau in "Resetting Xbox"

jsemrau — Tue, 07 Jul 2026 02:24:31 +0000

Over the last 8 years

Undead Labs has shipped nothing Compulsion shipped South of Midnight topping at ~1600 players on Steam Ninja Theory has shipped Bleeding Edge and Hellblade 2 and Double Fine shipped Keeper and Kiln.

It's a net positive that these studios and this culture is gone. I am sure that food was amazing.

New comment by jsemrau in "Zuckerberg says AI agent development going slower than expected"

jsemrau — Mon, 06 Jul 2026 23:41:32 +0000

Microsoft own Enterprise if you like it or not. And Microsoft will also own a large piece of the AI in Enterprise cake. https://www.microsoft.com/en-us/frontier-company

New comment by jsemrau in "Zuckerberg says AI agent development going slower than expected"

jsemrau — Mon, 06 Jul 2026 00:34:37 +0000

The last two years have been perfect for accumulating tech debt.

2023 you would have probably implemented your Agents with LangChain and RAG

2025 you'd use MCP and OpenAI/Anthropic Agent SDK.

2027 you will use a workspace frameworks (Amazon, Microsoft) sensor libraries and world models.

Agents are a fantastic generational technologies, but in mid-2026 the environment they are operating in is quickly changing.

The only way forward is to stay agile, understand model and vendor risk.

New comment by jsemrau in "Asian AI startups launch Mythos-like models"

jsemrau — Sun, 28 Jun 2026 05:31:10 +0000

The only way they can compete in Japan is the enterprise-game. Their partnership with Daiwa and MUFG is probably exactly what they should be doing. I doubt that they get that far though. Like Mistral who has partnerships with BNP Paribas and Airbus, they deploy in an on-premise or private cloud and in those settings, their models make good PoC's but cant compete where it gets interesting -- Workspace Agents. If you look at them from a Chatbot over RAG perspective, maybe they can do it. But that's tech from 2023.

New comment by jsemrau in "Asian AI startups launch Mythos-like models"

jsemrau — Sun, 28 Jun 2026 00:35:51 +0000

But did they deliver anything yet? There are many startups with notable founders that are not being able to get models launched. And even if they are able to launch, so they get PMF. In case of Sakana, they clearly focus on the Japanese market an have buildup a good pipeline on sovereign AI. But similar to Aleph Alpha or to a certain extend Mistral, I don't see how they can keep up.

New comment by jsemrau in "Algorithmic Monocultures in Hiring"

jsemrau — Tue, 23 Jun 2026 21:14:01 +0000

Interesting timing as Workday is facing Discrimination Claims in California doing the same thing.

https://www.yahoo.com/news/us/articles/california-judge-upho...

New comment by jsemrau in "Don't trust large context windows"

jsemrau — Sun, 14 Jun 2026 20:49:26 +0000

Just because you have a large garage doesn't mean you can park your car easily.

New comment by jsemrau in "Federal judge blocks H1B visa $100K fee"

jsemrau — Tue, 09 Jun 2026 05:23:35 +0000

> there was no qualified American applicants

I have been hiring for 20 years and find this increasingly impossible to believe. Please expand why that would be the case.

New comment by jsemrau in "Anthropic surpasses OpenAI to become most valuable AI startup"

jsemrau — Sat, 30 May 2026 15:20:15 +0000

I did a pair programming comparison over 3 month on Codex 5.2 and Claude Sonnet and my subjective experience was that based on cost and rollbacks to a previous commit Claude is significantly better. Especially in VS Code Copilot. I wrote a long Substack post about it. I would share its but its in the paywalled archive by now.

New comment by jsemrau in "Expertise in the age of AI"

jsemrau — Fri, 29 May 2026 19:35:46 +0000

In my opinion, the bread example doesn't really work that well because it bridges into the physical domain which most cognitive systems don't have access to. That said, for grounding context and therefore creating truth having a version of a world model is very important (See Yan LeCun's work). My experience is that given the right world to operate in, an agent can indeed find flaws in recipes and fix it even though the agent has not been prompted explicitly to do it. This world, as far as I understand it now, is a combination of sequential (at which step am I in a process), conversational (what was talked about alread/ what had I done already), and context memory (what is the frame or reference/plane of existence).

Self-correcting agents are already here: https://jdsemrau.substack.com/p/hyperagents-and-self-correct...

New comment by jsemrau in "I think Anthropic and OpenAI have found product-market fit"

jsemrau — Thu, 28 May 2026 04:56:38 +0000

Anthropic for sure. It's a useful professional product that I find many use-cases for in my professional and private life. OpenAI not so much.

New comment by jsemrau in "δ-mem: Efficient Online Memory for Large Language Models"

jsemrau — Sun, 17 May 2026 07:15:49 +0000

Thank you.

New comment by jsemrau in "δ-mem: Efficient Online Memory for Large Language Models"

jsemrau — Sun, 17 May 2026 00:04:58 +0000

That assumes one layer of memory. In my experience you need to have at least 4 layers of memory to work well. All of them have different requirements for retrieval. Everything that is in short-term memory (state of the app, current conversation, current workspace artefact) requires fast latency and precision. For example if you want to edit a segment in a financial analysis, a blog post, or a program you only want to edit this segment. RAG on a VectorDB is overkill in my opinion.

New comment by jsemrau in "δ-mem: Efficient Online Memory for Large Language Models"

jsemrau — Sat, 16 May 2026 13:08:46 +0000

I am currently working on deep context query which uses dynamically generated regex to pull only the relevant context blocks. By using lightweight RegEx pattern matching to detect semantic intent and filter structured context sections accordingly, you avoid the attention degradation that comes from stuffing semantically redundant information into the window

https://jdsemrau.substack.com/p/tokenmaxxing-and-optimizing-...