正在加载视频...
视频加载失败
Introducing Programmer You give it a task. It writes a program. The program runs. Try it:
20,890 次观看 • 8 个月前 •via X (Twitter)
11 条评论

How to compare it to Karpathy's autoresearch Programmer gives an AI agent a task and a set of real tools and an LLM. The agent writes a short program that orchestrates those tools, then the program actually executes. The agent decides what to search, what to fetch, what to ask, what to store, and when it's done. One prompt in, structured results out. Karpathy's autoresearch gives an AI agent a training script and a metric. The agent modifies the code, runs it, checks if the metric improved, keeps or discards the change, and repeats. The human writes the strategy; the agent is the tireless executor. The pattern is the same: the AI writes the code, the code runs, the results come back. Karpathy applied it to ML training while I have applied it to general tasks with live data sources. What's the same: - The human describes intent, not implementation. - The AI writes real, exeutable code - not tool calls or JSON. - The code runs against real infrastructure - not a simulation. - Every execution produces a trajectory that can train the next version of the model. What's different: - Autoresearch modifies existing code. Programmer generates code from scratch. - Autoresearch optimizes a single metric. Programmer synthesizes information from multiple sources. - Autoresearch is a loop (modify → run → eval → repeat). Programmer is a single pass (generate → execute → converge), though the architecture supports iteration. cc @karpathy

a denser task for the Programmer agent: I'm visiting Tokyo for 3 days. Find the best ramen in Shinjuku, sushi near Shibuya, and give me historical context on Japanese food culture

best joke project, it does not even know what's claude opus4.6

when i want to create a dasboard like this below to create a construction dashboard in Figma, which will later be implemented in Microsoft Power BI and expanded into a SaaS web application. We will start with one dashboard page as a trial. If the delivery meets expectations, we will continue with the remaining 4 pages. Project scope: •Total: 5 dashboard pages •First page: 20% of total budget (trial page) •Remaining 4 pages: 80% after approval of the first page First page requirements: •KPI Cards: Progress %, SPI, CPI •Charts: Planned vs Actual (S-Curve), Progress by Zone •Filters: Project, Zone, Date Design expectations: •Clean and professional layout •Suitable for Power BI implementation (simple and practical UI) •Scalable for multiple pages (SaaS-ready design system) •Organized components (colors, fonts, styles) Deliverables: •Fully editable Figma file •Organized layers and components •Proper spacing, alignment, and hierarchy it only accepts until 500 words

I thought we already had that and it was called AI

love your experiment w/ programmer any plans to add evals to it? when you explain the single pass choice, you mention "converge" - isn't this an iterative process?

Amazing work over the past two weeks. Wondering how I can leverage some/all of these apps inside a restricted set of actors such as a class (collection of students and agents), College(collection of students, faculty, staff, and agents) and such. made this recently-

@HyperspaceAI Cool

@HyperspaceAI wallet show returns 'install ethers' on v5. 1.6 SEA binary on macOS arm64, even with - - earn flag

Cool stuff! Is there a github for it?

@HyperspaceAI @grok explained use case


