Yann Leflour

LLMs can't synthesize information and nobody's talking about it I was looking into Yoji Akao's book on Quality Function Deployment last week. 1st thing I noticed: half the pages are charts Guess what it takes for an LLM to convey the same level of information as a single one of these charts?

Ethan Mollick@emollick.bsky.social

An OpenAI researcher (Noam Brown) was generally mocked last month for suggesting that even air-gapped AI models on separate computers would find a way to communicate with each other, perhaps through CPU temperature changes. Given this latest report from OpenAI, it doesn’t seem so implausible.

New EP dropped. Love this band

While we keep trashing on slop code, let's remember that this is the experience plenty users still have to deal with Sure LLMs can't do everything, but today's models are likely to spot these issues & display proper errors without extra instructions

Bruce Lawson@brucel.bsky.social

I have found the Worst Website in the World. It is the official Indian Railways ticket booking site. After the 96 step sign up process, it randomly throws you out; it requires repeated entry of information it already has, and occasionally throws useless error messages. Wanted 2 seats, now have 4.

That was my main pain point of 2024. I kept developping custom tools to fill in the gap and try to consolidate into a product Every new model/harness update filled part of the gap and made my tools obsolete I quit trying to improve the "code writing" part around that time

dax@thdxr.com

a weird inversion with LLMs is the models improve faster than the tinkerers when i see people with custom workflows and setups they're all addressing problems that don't exist anymore the person naively using vanilla codex is more likely to be experiencing state of the art

Now my top priority is - Making legacy codebases AI ready through reverse engineering pipelines - Keeping the overall product legible to the product teams with graphs, schemas, ... as most teams struggle at keeping track of the changes This part hasn't changed since GPT 3 release

16 July

Oh I see. I should probably refresh my feed for news on the Kimi 3 release

Anthropic {bot}@anthropicai.xmirror.bot

We've reset 5-hour and weekly rate limits for all users.

We've taught these models to write, to code, to reason out loud. We haven't taught them to draw, to take a mess of relationships and compress it into one thing you can see Until we do, "synthesis" stays a human skill. And I'm not sure the labs are even pointing at it

Be honest. When did an LLM last hand you a diagram you actually used without redrawing it?

I draw to explain concepts, problems, solutions to my team almost every day A freeform sketch settles an argument faster than any documentation I've written. I'll go further: knowing how to draw is a core engineering skill, and I think we should teach it the way we teach coding

It's also exactly what LLMs can't do. Ask one for a diagram and you'll see it. It produces a picture of a chart, not a chart that does the thinking for you So here's what I keep coming back to

I ran the numbers. For one A3-size chart, the LLM wrote down: - 8 sections - 33 paragraphs - 1,048 words - 7 A4 pages for the same level of information

A person can scan the chart in about 30 seconds and you can kickstart a team discussion around the chart in a matter of minutes Meanwhile, the LLM-produced document requires the entire team to go through the 7 pages and mentally map the related information before anything

github.comGitHub - obra/superpowers: An agentic skills framework & software development methodology that works.An agentic skills framework & software development methodology that works. - obra/superpowers

12 Feb

Superpowers is an amazing piece of engineering It may seem like a simple brainstorm / plan / execute worflow, but the execution is top notch If you care about software craftsmanship, then it's a must have in your toolset The output's quality for each step is unparalled github.com/obra/superp...

17 Oct 2025

3. I've invested in learning how to use this tool in my field of expertise. Holy cow, it's transformational but will be a plague for human intent, knowledge, and creativity.

Justin Wolfers@justinwolfers.bsky.social

There are basically only two positions in the debate about AI. 1. I’ve barely invested any time in learning how to use it effectively. AI sucks. 2. I’ve invested in learning how to use this tool. Holy cow, it’s transformational. Position #1 has lower barriers to entry.

29 July 2025

It's July 2025 and prompting is still a skill A skill I have a hard time teaching Some people just get it, but most struggle And no, LLMs are not very good at prompting themselves without another prompt to guide them

28 July 2025

If you are tired as me of working with native path and fs libraries in NodeJS, you might like my www.npmjs.com/package/@sy... NPM package. A library to finally turn verbose JS filesystem code into elegant one-liners

25 July 2025

Cursor has a system prompt issue. I'm using LLMs for a complex software migrations and have leveraged everything I know about prompting. While the model is the same as Claude Code, Cursor seems unable to follow the provided instructions. Cursor simply feels broken now

23 July 2025

LLMs have helped me build dozens of libraries I've always wanted in Javascript My favorite one is www.npmjs.com/package/@sy... The concept is simple, what if you could write clean template strings in JS ? I've been using it for clean error formating, LLM messages, and more

7 Apr 2025

When I'm hired to enable a team with AI tools what I'm actually paid for is asking the right questions to reverse engineer how the team works When I have that, building the tools is the easy part Also why I fear that the agents trend will do a lot of harm to craftsmanship

31 Mar 2025

If you're building LLM based products, check out napkin.ai Blending pre-made templates with LLM formatting may seem like a straightforward idea but they've executed it really well with a clean UX And if you need to fill a presentation with corporate BS, just ask it to generate the text as well 😅

youtube.comVibe Coding with AI in 2025YouTube video by Programmers are also human

18 Mar 2025

I've used vibe coding (AI codes without you reviewing) to build a visualizer for a graph with Cline Would never trust a core product or mission critical system coded like this People are mistaking vibe coding and vibe engineering youtube.com/shorts/1_rSr...

7 Mar 2025

An upside of LLMs is that, just by using them, people will get better at providing instructions and the right context The downside is that they'll likely stay to get frustrated when you take over 1mn to respond

6 Mar 2025

While illustrating the many disciplines connected to AI and LLMs I realized that I don't have a term for UX Designers / PMs focused on AI Really shows how under-explored UX for AI is I agree it's hard, but we could be doing so much more than chats

4 Mar 2025

Has anyone studied the impact of image and sound on multi-modal LLMs ? Like if you're working with the LLM, does passing it classical music unrelated to the task influences its response ? What about art pieces ?

20 Feb 2025

My son (5 yo) is fighting with his cousins (6 yo) about the right #Uno rules Some things don't change...

harper.blogMy LLM codegen workflow atmA detailed walkthrough of my current workflow for using LLms to build software, from brainstorming through planning and execution.

20 Feb 2025

Wow, @harper.lol's workflow is incredibly efficient with gemini 2 flash thinking and its massive context I used to build custom backends to iterate on ideas step by step. Now I just need the LLM We're truly shifting from coding LLMs to engineering ones harper.blog/2025/02/16/m...

youtube.comFabrice Bernhard & Yann Leflour - Swift to React Native: AI-Driven Transpilation | RNLConf 2024YouTube video by Theodo UK

18 Feb 2025

Finally! Our React Native London talk with @fab-ber.bsky.social on AI assisted migrations is available online Discover how we achieved a 9 months migration in just 4 months by combining solid engineering expertise with a custom AI workflow www.youtube.com/watch?v=sz_S...

30 Jan 2025

This also means that Deepseek R1 will be a potential attack vector. Willingly put inside a state system Imagine if they trained the model to act maliciously when it figures out that it's being used inside a targeted infrastructure I wonder what it could realistically achieve

Michael Kevin Spencer@aisupremacy.bsky.social

India lauds Chinese AI lab DeepSeek, plans to host its models on local servers. The approval to allow DeepSeek to be hosted in India appears contingent on the platform storing and processing all Indian users’ data domestically, in line with India’s strict data localization requirements.

Yann Leflour reposted

tldraw

@tldraw.com

28 Jan 2025

you can now secretly sign in at tldraw.com/preview

23 Jan 2025

"Gemini 2 Flash Thinking" is a very impressive model I'm having discussions with it about high level and abstract concepts and it just "gets it" The 1M tokens context also means that you can have extremely long conversations with high retrieval rates and without running out of chat history

aistudio.google.comGoogle AI StudioGoogle AI Studio is the fastest way to start building with Gemini, our next generation family of multimodal generative AI models.

23 Jan 2025

You should try it on aistudio.google.com This type of model will likely be my go to for high level agentic systems

21 Jan 2025

Quick PSA to every developer applying for their first job: It's time to include LLM tools in your resume (Copilot, Cursor, ...) along with any expirement you've done with them I'm currently reviewing resumes and none mention these, so it's still a good way to set your profile appart

21 Jan 2025

I'm also asking candidates to provide me with links to some chats with ChatGPT (or equivalent) in order to assess how they work with LLMs My guess is that this will become a common practice in a near future

7 Jan 2025

I stopped my #devin (yes the "AI Engineer") subscription, felt too uneasy about merging those pull requests It was a fun (though way too expensive) trial It's hard to put it, but to me it's like most "AI art", the demo is impressive but it clearly lacks any intent

7 Jan 2025

And I believe intent is crucial when developping. I'd rather have a human one I can train and steer than this husk of a developper Copilots and specific agents are not the same, they're like advanced paintbrush, helping you with the first draft or tiny touches. So I'll stick with those for now

@vercel.com's AI SDK is amazing to work with But I don't get why I had to implement a middleware to add the last message from the prompt to the result if its role is "assistant" Should be a default behavior Anyway, if you need it it's "includeAssistantMessage" in @synstack/llm/middleware

npmjs.comnpm: @synstack/llmImmutable, chainable, and type-safe wrapper of Vercel's AI SDK. Latest version: 2.1.0, last published: an hour ago. Start using @synstack/llm in your project by running `npm i @synstack/llm`. There is 1 other project in the npm registry using @synstack/llm.

Works with streaming as well by the way www.npmjs.com/package/@sy...

Even though I would not be surprised to see something like this implemented, there is no way it's true But seeing how this is spreading, it shows that we're far from achieving widespread understanding of what AI really is

This post is unavailable.

github.comsynscript/packages/llm at main · pAIrprogio/synscriptLibraries for immutable and chainable LLM compatible scripting - pAIrprogio/synscript

Finally released @synstack/synscript v2 with the #llm package migrated to @vercel.com's AI SDK 🎉 The Llm library is now a chainable, immutable, and type-safe wrapper over the ai sdk If you want to improve your prompting DevX in Typescript, look no further github.com/pAIrprogio/...

4️⃣ Prompt chaining Because prompt scripting is mostly async, I disliked writing await and Promise.all to whenever working on message arrays So, the message() param can be a Promise<Array<Message>> or an Array<Promise<Message>>, making it simpler and cleaner to build out your prompt and chain agents

There are more libraries I've built @ www.npmjs.com/org/synstack Make sure to check out "@synstack/fs" which the other one I'm most proud of I've been using them over the last 6 months to help teams succesfully triple their velocity during software migrations and they'll keep on getting better

github.comsynscript/packages/llm at main · pAIrprogio/synscriptLibraries for immutable and chainable LLM compatible scripting - pAIrprogio/synscript

2️⃣ Type safety There's type safety at every level: 🔶 Tool definitions, filters, & outputs 🔶 Calling the llm with missing required parameters will trigger a TS error 🔶 Message template string values are type-safe as well Even my test coverage is mostly type testing for now github.com/pAIrprogio/...

github.comsynscript/packages/llm/src/completion.builder.test.ts at main · pAIrprogio/synscriptLibraries for immutable and chainable LLM compatible scripting - pAIrprogio/synscript

3️⃣ Template strings I hate the base JS template string because it breaks indentation So I wrote a separate library to fix that. 🔶 Trims text and removes extra indentation 🔶 Auto-resolves promises 🔶 Arrays to separate lines 🔶 Propagates nested indentation github.com/pAIrprogio/...

github.comsynscript/packages/llm at main · pAIrprogio/synscriptLibraries for immutable and chainable LLM compatible scripting - pAIrprogio/synscript

This allows me to benefit from the countless integrations and tons of contributions to AI SDK while keeping my target level of developer experience Here's a quick tour of some of the features:

github.comGitHub - ladjs/superagent: Ajax for Node.js and browsers (JS HTTP client). Maintained for @forwardemail, @ladjs, @spamscanner, @breejs, @cabinjs, and @lassjs.Ajax for Node.js and browsers (JS HTTP client). Maintained for @forwardemail, @ladjs, @spamscanner, @breejs, @cabinjs, and @lassjs. - ladjs/superagent

1️⃣ Immutability + Chaining I've always been a fan of superagent's clean immutable and chainable API allowing me to build upon a base instance of query github.com/ladjs/super... This was my target to replicate, but for LLMs and most of my libraries published under @synstack operate that way

github.comsynscript/packages/llm at main · pAIrprogio/synscriptLibraries for immutable and chainable LLM compatible scripting - pAIrprogio/synscript