devnullg

Former AI skeptic → now making money with it
Sharing working workflows and prompts
No fluff. Just what works in 2026

[devnullg @devnullg]

@dair_ai Which of the three trips people up most in practice - the terminal-only feedback ambiguity, or the memorization instead of generalizing?

[devnullg @devnullg]

@omarsar0 Running a harness against a live recommender at billion-user scale is wild - what's it actually enforcing: hard action guardrails, or more a ranking policy the agent can still nudge?

[elvis @omarsar0]

Massive paper from Meta.

I like this one because it shows the use of agent harnesses for production-grade recommender systems.

Details below:

This is one of the more convincing agent deployments I've seen.

It runs against a live production recommender serving billions of people and reports A/B results.

Sustaining a recommender is continual optimization work. Content shifts, user behavior shifts, upstream models shift, and the choices governing retrieval, ranking and serving have to be revisited.

Human engineers test those changes through online experiments, which is slow enough that parts of the system go unrevised.

In CORAL, each cycle the agent observes operating signals, reasons over a memory of past decisions and their measured outcomes, and invokes tools including a numerical optimizer that keeps every change inside a fixed operating budget.

The policy improves in context from its own prior actions, with no parameter updates.

Across two large social platforms, the same harness improves engagement at no additional serving cost on one and reduces serving cost without degrading engagement on the other.

Performance improves as the loop iterates.

The guardrail design carries as much weight as the agent. A bounded change budget makes this safe to run against production.

Paper: https://t.co/G46EgVuPMR

Chat with Paper: https://t.co/KlYFT8dAFD

[devnullg @devnullg]

@omarsar0 Outliving the host is the part that gets me - most agent state I've seen dies with the process. What actually persists across the restart - just memory, or the in-flight task state too?

[devnullg @devnullg]

@QwenDevs How does it decide when to fall back from semantic to BM25/rg for a query - automatic based on query shape, or does the agent have to pick the mode explicitly?

[Qwen Developers @QwenDevs]

Alibaba’s Zvec team open-sourced zg, a local search tool for developers and AI agents.

• Local-first
• Works out of the box with popular agents
• Semantic, BM25, hybrid, and rg search in one tool

Why they built it and how it works: https://t.co/HzihCJh5l8

[devnullg @devnullg]

I wasted $23 on API calls last month doing this:

Testing prompts directly in production scripts.

Now: draft in playground, count tokens, THEN wire it up. Same results, a third of the cost.

Learn from my bill.

[devnullg @devnullg]

A 50-year-old grand challenge fell in a single competition cycle - and the winners gave away all 200 million answers for free.

November 2020, CASP14. AlphaFold2 predicts protein structures with accuracy rivaling slow, expensive lab methods.

The problem dates back decades: proteins fold into shapes that determine what they do, and predicting shape from sequence was biology's holy grail. Christian Anfinsen argued back in his 1972 Nobel lecture that the sequence alone should contain the answer.

DeepMind delivered it. Hassabis and Jumper got the 2024 Nobel in Chemistry for it. Then they released the AlphaFold Database with 200M+ structures, free for everyone.

Open-sourcing the result multiplied its scientific impact instead of losing money.

What other grand challenges are waiting for the same treatment?

[devnullg @devnullg]

Genuine question for the experienced folks:

At what point did AI go from 'fun experiment' to 'part of how I work'?

Trying to figure out if there's a moment or if it just... happens.

[alphaXiv @askalphaxiv]

“Fast Weight Attention for Continual Learning”

Fast-weight models try to make attention cheaper by continuously rewriting a small fixed-size memory.

This paper shows that this rewrite should behave more like online learning from what the model just predicted to what actually came next.

So they made adaptive memory updates that decide how much to learn, forget, and rehearse, while keeping constant-size memory and efficient training.

https://t.co/aG0aFJp2SL

[devnullg @devnullg]

@omarsar0 two-hop subgraph plus a fixed candidate budget sounds like the real tradeoff. Does the win come from the graph structure itself, or mostly from capping how much gets pulled at retrieval time?

[devnullg @devnullg]

@rohanpaul_ai generating the harness on demand instead of picking one upfront is such a different mental model. Does it re-generate the harness mid-task if the situation changes, or is it locked in once the run starts?

[devnullg @devnullg]

Tested 5 AI meeting note-takers this month.

Kept: 1 (the ugliest interface, best summaries)
Deleted: 4

Lesson: features don't retain users. Accuracy does.

Which one did you settle on?

[devnullg @devnullg]

@Ezekiel_dev1 That's the exact failure mode I'd expect - garbage data silently poisoning a CRM is way worse than a slow save. 20 min sounds cheap for that insurance.

[devnullg @devnullg]

Set up an n8n workflow that drafts replies to common emails.

Setup: 45 minutes.
Saves: ~2 hours/week.
Cost: $0 (self-hosted).

Automation isn't about replacing yourself. It's buying back evenings.

What did you automate lately? #automation

[elvis @omarsar0]

Nice paper showing just how far you can push an agent harness.

In most setups, the default coding agent harness is static. Capabilities get wired in at design time, and the run has no way to change how it is being executed.

openJiuwen is an open-source harness built to fix that.

It reaches 82.6% on SWE-bench Verified and 87.19% on Terminal-Bench 2.1, ahead of the strongest official leaderboard entries by 3.4 and 3.39 points.

A bit on how openJiuwen works:

Rail-based composition lets developers assemble single agents, delegated sub-agents and swarm flows over one shared execution substrate.

Runtime evidence from semantic diagnostics, execution outcomes and task progress then reshapes context, feedback and task control while the run is still going.

The model policy stays fixed throughout, so the gains are attributable to the harness rather than to the model underneath it.

Paper: https://t.co/QegawA8cxB

Chat with Paper: https://t.co/DTmu2Dx8vu

[devnullg @devnullg]

@omarsar0 Rail-based composition is the part I keep coming back to - assembling agents, sub-agents, and swarms on one substrate sounds like it could get messy fast. Any tooling to trace which rail caused a regression?

[devnullg @devnullg]

@rohanpaul_ai The remember-wins-and-failures part is what gets me - most agent setups I've seen just retry blind after a failed branch. Does the tree search actually prune dead branches, or just deprioritize them if compute allows?

[devnullg @devnullg]

In 2020, OpenAI published something close to a price list for intelligence.

That chart is why GPU spending went nuclear.

"Scaling Laws for Neural Language Models", January 2020. The finding: language model loss falls as a smooth power law as you grow parameters, data and compute - predictably, across seven orders of magnitude.

Not breakthroughs. Not tricks. A curve you can extrapolate before spending the money.

Five months later came GPT-3 at 175 billion parameters - a bet placed directly on that curve. It paid.

Then DeepMind's Chinchilla paper (2022) corrected the ratio - most models were starved of data relative to size - and the industry re-planned its training runs overnight.

Intelligence had become an engineering budget line.

Nobody agreed to that out loud. A chart did it.

[devnullg @devnullg]

My AI agent's weekly report would say:

Tasks completed: 34
Times I interrupted it mid-run: 11
Times it was right anyway: 10

Trust issues work both ways, apparently.

[devnullg @devnullg]

Real talk on AI costs for beginners:

Month 1 I spent: $47 on APIs
Value created: ~$300 in saved time + $127 earned

You don't need a big budget. You need to track every dollar like it's your last. #AIIncome