Guillaume Loret retweeted
Can DeepSeek V4.1 Flash match Opus 5.5 on real knowledge work?
I ran both on 50 of my anonymized Granola calls, turning them into one action log.
Guillaume Loret retweeted
Today we're releasing WorkBot and OpenWork in Slack.
WorkBot is a simple chat interface for people who don't know or care what a skill or an MCP server is. They just want to get their work done.
You set up skills, connectors, and models once, from any MCP client. They're then available everywhere your team already works: Slack, WorkBot, the OpenWork app, and even Cowork and Codex.
OpenWork is free for up to 5 users and works with any model provider.
Guillaume Loret retweeted
Biggest misconceptions bout OpenWork is that we're a desktop app.
OpenWork is a way to deploy AI capabilities safely to your team.
Whether it's MCPs, AI Models, or skills - admins can configure them all in OpenWork and distribute them across a wide way of surfaces.
Different harnesses, slack, simple chat interfaces, or the desktop app are all supported.
Guillaume Loret retweeted
In a world where you can do 100 parallel tasks remember that same world allows *anyone* to do the same.
It is not ambitious to take on more work if your work is easy.
But if anyone can have 100 parallel tasks going on, why aren't there more people getting ahead?
The reason is that is in a world where you can move faster, moving faster in the correct direction matters even more. You can compound faster, create more distance between you and everyone else.
The important things haven't changed: deciding what to work on, sticking to it, and focusing.
Your software factory isn't creating the gap you think it does. The gap you should care about is the gap in market value you captured, not the code you were able to produce.
You are not working in isolation, you're competing against the entire world.
Someone doing a "software factory" is not a scarce resource. What is scarce is attention and discernment. Intelligence being "on tap" doesn't change that.
The truly ambitious don't only take on more work. They take on a particular kind of work in a particular way.
Guillaume Loret retweeted
we made our terrible onboarding, a bit less terrible
(jk we think it's great now, but do let us know)
Guillaume Loret retweeted
Why GPT-5.4, Claude, and Gemini can't agree on basic, real-world facts thenewstack.io/frontier-llm-…
Guillaume Loret retweeted
Android Bench evaluates LLM coding performance by focusing on the nuances and platform expertise that are unique to Android.
See how your favorite model stacks up → goo.gle/4swdpRe
Are the results what you expected? 👀 Let us know in the replies!
Guillaume Loret retweeted
Wait, there is also an English version: correctiv.org/en/europe/2026…
Guillaume Loret retweeted
Ce visuel de Bloomberg est excellent (et un peu terrifiant aussi)...
C'est l'explication de ce qu'est une bulle et là il s'agit d'une bulle dans l'IA. Quand les investissements deviennent circulaires, sans irriguer l'économie réelle
- OpenAI achète des puces Nvidia
- Nvidia investit dans OpenAI
- Oracle loue des puces Nvidia à OpenAI
- OpenAI prend des participations dans AMD
Fasten your seatbealts, it gonna be shaky.
On rappellera que la somme totale des engagements pris par OpenAI se chiffrent à 1000 milliards de dollars.
Guillaume Loret retweeted
Someone at a16z claimed a few weeks ago that 80% of Bay Area startups are building on Chinese open source models. The graphic below shows Chinese model downloads exceeding US models on HuggingFace.
Guillaume Loret retweeted
60% of CS PhDs in the US are not American.
53% of CS Masters grads in the US are not American.
BigTech cos have <15% employees on H-1B.
BigTech hiring isn't necessarily biased against Americans; most Americans just don't study advanced CS!
Guillaume Loret retweeted
Why can AIs code for 1h but not 10h?
A simple explanation: if there's a 10% chance of error per 10min step (say), the success rate is:
1h: 53%
4h: 8%
10h: 0.002%
@tobyordoxford has tested this 'constant error rate' theory and shown it's a good fit for the data
chance of success declines exponentially
Guillaume Loret retweeted
UTF-8 🤦♂️
I already knew about the "confusables", e.g.: e vs. е. Which look ~same but are different.
But you can also smuggle arbitrary byte streams in any character via "variation selectors". So this emoji: 😀󠅧󠅕󠄐󠅑󠅢󠅕󠄐󠅓󠅟󠅟󠅛󠅕󠅔 is 53 tokens. Yay
paulbutler.org/2025/smugglin…
Guillaume Loret retweeted
Replying to @karpathy
This is possible via Unicode Tag code points, read and write hidden text (ASCII Smuggling). No need for tool use, it's quite amazing.
Allow-listing tokens is an important mitigation when building LLM apps
Claude and Grok are still vulnerable I believe
embracethered.com/blog/posts…
Guillaume Loret retweeted
People have too inflated sense of what it means to "ask an AI" about something. The AI are language models trained basically by imitation on data from human labelers. Instead of the mysticism of "asking an AI", think of it more as "asking the average data labeler" on the internet.
Few caveats apply because e.g. in many domains (e.g. code, math, creative writing) the companies hire skilled data labelers (so think of it as asking them instead), and this is not 100% true when reinforcement learning is involved, though I have an earlier rant on how RLHF is just barely RL, and "actual RL" is still too early and/or constrained to domains that offer easy reward functions (math etc.).
But roughly speaking (and today), you're not asking some magical AI. You're asking a human data labeler. Whose average essence was lossily distilled into statistical token tumblers that are LLMs. This can still be super useful ofc ourse. Post triggered by someone suggesting we ask an AI how to run the government etc. TLDR you're not asking an AI, you're asking some mashup spirit of its average data labeler.
Guillaume Loret retweeted
The PocketPal AI mobile app (iOS and Android) was open sourced yesterday. The project looks fun and seems like a useful playground for LLMs at the edge. Looking forward to where the community takes it from here.
github.com/a-ghorbani/pocket…
Shoutout to the author @ghorbani_asghar
Guillaume Loret retweeted
Introducing Meta Segment Anything Model 2 (SAM 2) — the first unified model for real-time, promptable object segmentation in images & videos.
SAM 2 is available today under Apache 2.0 so that anyone can use it to build their own experiences
Details ➡️ go.fb.me/p749s5
Guillaume Loret retweeted
Project management is like a puzzle. Different techniques fit different circumstances. It's always good to have a big box to make sure you can tackle any situation, especially for dynamic startup projects.
forbes.com/sites/abdoriani/2…






















