Alberto 思辰

Today I am leaving Vietnam 🇻🇳 Food: 8.5/10 Landscapes: 9/10 Accommodations: 7/10 People: 8/10 Coffee: 10/10 Best place to stay considering all factors: Da Nang for food: Hanoi For countryside and relaxing: - Tam Coc (Ninh Binh) - Da Lat I almost always eat local food, but it’s a bit more expensive than Thailand 🇹🇭, but quantity of food is greater. Average price for a meal: 50-100k dong (2-4$) tea is always free! If you like soups and not to spicy food is better than Thailand, but I missed so much the pad see ew! To travel from north to south or viceversa I suggest you to take the night cabin buses. Monthly budget: 1K $, but I spent a little bit less You can easily stay on 800$ or less if you don’t travel a lot and you stay in the same place! Driving a motorbike is a bit crazy, but in cities like Da Nang roads are larger and there is less traffic! Heading to Kuala Lumpur now! #buildinpublic #digitalnomad

@ardent__dev get distribution without ads or seo

What's more valuable skill today? - Coding skills - Marketing skills

@omarvvvr Yes, tried for fun doing some task the old way again, felt amazing but it took me x5 the llm time.

Be Honest: Can you CODE without AI...?

I recently cancelled the github copilot plan, but there was a feature I was using a lot: auto commit generation. Since I want to have control on commits to keep a clear history I created a vscode extension to do exactly this. you can decide what llm provioder to use (even a local model). github.com/carbogninalber…

@BlockedPaths @william_arin Me too specifically for tool calling and other parsing issues i found.

@william_arin @doregex I believe dario or elon said they are bout 3 months behind. When we release a new model they distill their modes with it, Kinda cheating. Ive built a harness with deepseek.

Kimi K2.7 did not impress me in the slightest. No where near 5.5 or deepseek.

After a couple of months of experimenting I kinda find a way to use them for 90% of coding tasks, even complex one. I don't know, but I find myself writing detailed prompts and create plans before implementing and in particular using subagents so that the main llm context window stays under 100k-150k tokens, above that chinese models in general starts to get confused (the attention is not where I need it to be) I use them mostly on existing codebases though, so maybe on new projects they might create worse code if they are not thoughtful of all the edge cases. One factor is that they achieve good results by thinking a lot, but the trend seems leaning towards optimization right now, see kimi2.7 code

@doregex @BlockedPaths Yes, but I'm not sure if they will catch up that quickly. Last december it felt like Chinese models lagged 4-5 months behind. But GPT 5.0 in september already understood intent very well, and 9 months later Chinese models are still not at this level I think.

用 OpenCode 加这个插件，换成 DeepSeek: const DEEPSEEK_PROVIDER = "deepseek" const ARRAY_FIELD_HINTS = new Set([ "args", "arguments", "extensions", "files", "patterns", "paths", ]) function modelRef(model, provider) { const providerID = provider?.info?.id ?? provider?.info?.name ?? provider?.id ?? provider?.name const modelID = model?.id ?? model?.modelID ?? model?.name return `${providerID ?? ""}/${modelID ?? ""}`.toLowerCase() } function isDeepSeek(model, provider) { return modelRef(model, provider).includes(DEEPSEEK_PROVIDER) } function looksLikeArray(value) { const trimmed = value.trim() return trimmed.startsWith("[") && trimmed.endsWith("]") } function fieldLooksArrayLike(key) { if (!key) return false const lower = String(key).toLowerCase() return ARRAY_FIELD_HINTS.has(lower) || lower.endsWith("list") || lower.endsWith("array") } function fieldLooksPathLike(key) { if (!key) return false const lower = String(key).toLowerCase() return lower === "path" || lower === "filepath" || lower === "absolutepath" || lower.endsWith("path") || lower.endsWith("file") } function unwrapDegenerateMarkdownLink(value) { return value.replace(/\[([^\]]+)]\(https?:\/\/([^)]+)\)/g, (match, text, url) => { const urlWithoutProtocol = url.replace(/\s+/g, "") return text === urlWithoutProtocol ? text : match }) } function repairValue(value, key) { if (value === null) return undefined if (typeof value === "string") { if (fieldLooksPathLike(key)) return unwrapDegenerateMarkdownLink(value) if (fieldLooksArrayLike(key) && looksLikeArray(value)) { try { const parsed = JSON.parse(value) if (Array.isArray(parsed)) return parsed } catch {} } if (fieldLooksArrayLike(key) && value.length > 0) return [value] } if (Array.isArray(value)) return value.map((item) => repairValue(item)) if (typeof value === "object" && value) { if (fieldLooksArrayLike(key) && Object.keys(value).length === 0) return [] const repaired = {} for (const [childKey, childValue] of Object.entries(value)) { const next = repairValue(childValue, childKey) if (next !== undefined) repaired[childKey] = next } return repaired } return value } function repairArgs(args) { if (typeof args !== "object" || !args) return args return repairValue(args) } export default async function deepseekToolHarness() { const deepseekSessions = new Set() return { async "chat.params"(input) { if (isDeepSeek(input.model, input.provider)) deepseekSessions.add(input.sessionID) else deepseekSessions.delete(input.sessionID) }, async "experimental.chat.system.transform"(input, output) { if (!isDeepSeek(input.model)) return output.system.push([ "DeepSeek tool harness:", "- Tool arguments are consumed by programmatic APIs, not rendered as chat markdown.", "- Omit optional fields instead of sending null.", "- Send arrays as JSON arrays, not stringified arrays or bare strings.", "- Send file paths as plain strings; do not markdown-link path values.", ].join("\n")) }, async "tool.execute.before"(input, output) { if (!deepseekSessions.has(input.sessionID)) return const before = JSON.stringify(output.args) output.args = repairArgs(output.args) const after = JSON.stringify(output.args) if (before !== after) console.info(`tool_input_repaired:${input.tool}`) }, } }

@Gracker_Gao 这是温度的问题，等等他们修一下吧

GLM5.2 ......你这让我如何放心啊，Claude Code 没这问题

@DanielSmidstrup it's all about let knowing the llm what code to write.

@DanielSmidstrup 90%? Maybe close to 99%

At this point if AI writes 90% of code, who even survives in tech??

@william_arin @BlockedPaths I mean, do you remember when models ask about the quality of the response or what you would change? that's the part that chinese models miss.

Yes, but that comes at a cost of having gathered many months of user interaction to finetune the model in agent chat alignment, or only god know what is behind after you sent your prompt. I think the "magic" is either some part of the model trained to expand your prompt or another model doing that job. Who knows?

No way you get a SOTA level of intelligence by selfhosting in a 6k hardware, you need like 150k$ server at least. You can run qwen 3.6 35B A3B on a consumer hardware at great speed quant 4bit or even 8bit smoothly, but you need to threat this agent as an implementer. By using a small amount of token from a sota model like gpt5.5 you can prepare detailed plans and give little room for small decision to the small models. There is this project that is promising btw: github.com/antirez/ds4

The Fable/Mythos news has me wondering how much AI access I should actually own locally. If a beginner had ~$6k to get into local models, what would you tell them to buy? Mac Studio? Mac mini? RTX 5090 build? Used dual 3090 rig?

@opencode please introduce a go 30$ subscription plan 🙏

@AlickZheng157 @buildwithhassan 我通常先让 GPT-5.5 做计划，然后让 DeepSeek 按这个计划工作，最后再让 GPT-5.5 看一看 DeepSeek 做得怎么样

@doregex @buildwithhassan deepseek V4 pro 高推理模式，幻觉严重。

my personal chinese coding model tier list from daily usage: tier 1: GLM-5 ≈ GLM-5.1 ≈ Kimi K2.6 > Qwen 3.7 Max (somehow the most expensive and not the best) tier 2: Qwen 3.7 Plus (under 200K context) > DeepSeek V4 Pro ≈ V4 Flash ≈ MiMo 2.5 Pro not there yet: MiniMax, Hunyuan testing K2.7 Code soon and GLM-5.2. ranking might change by morning.

@buildwithhassan it's already present in opencode go plan

kimi k2.7 code is so new that opencode doesn't have it yet

@nalinrajput23 Deepseek is so cheap, you prepare the plan via gpt and let deepseek implement it. Currently I am trying to use kimi2.7 instead of gpt, let's see

as a dev, which is the best Open Source AI model?

@sflorimm Building a linear app alternative fully opensource (apache 2.0) self-hostable with builtin agents. Now working on having task assigned to agents running in containers with opencode and your fav model

I want to connect with more founders builders vibe coders AI enthusiasts UI designers If you’re someone building with AI right now, drop it in the comments and let's connect

Introducing the Fusion API, the smartest compound model in the market. Fusion achieves Fable-level intelligence at half the price. How it works 👇

@aditiitwt Overall (cost/intelligence)? I think deepseek v4 pro at max. The only thing you have to keep in mind is that if you are specific enough and prepare plan before implementing the code the difference from using closed SOTA models in negligible on standard tasks.

Guyys be honest, which is the best open source AI model? 1. Qwen 3.6 2. DeepSeek 3. GLM-5 4. Kimi K2.6