Machine learning and Indie hacking.
London
Joined August 2012
- Tweets902
- Following1.9K
- Followers30
- Likes2.3K
Yann Lamidon retweeted
A 9B model could reach GPT-5 / Opus-4.5-level reasoning 🧠, nearly free ⚡🆓, on your own desktop 💻.
No additional post-training. Much less jagged 🧩 generalization.
This was my intern Panagiotis’ summer project. Still a long way to go on engineering, but I strongly believe this direction will keep getting better.
Small models may have a lot more intelligence 🧠 inside them than we think.
arxiv.org/abs/2609.38104
Yann Lamidon retweeted
Lots of founders are crashing out
A year ago AI felt like an incredible opportunity.
Most of us were able to ship faster, remove admin, and get time back.
Lots of engineering founders were even able to explore marketing for the first time: SEO, organic social, outbound.
Building complex systems that felt extremely productive.
But since then, a lot has happened.
Here’s some of the themes I’m seeing:
AI marketing isn’t working. At first it feels impressive and it looks good but for most it’s not driving results. This is made worse by the fact everyone has access to the same tools. Inboxes are flooded, buyers are fatigued, the economy is flat.
Engineers are fried. Yes you can ship more but for many the flow state has gone. It’s a new way of working and it’s not for everyone. The speed leaves many of us exhausted before 11am.
There’s a feeling of what’s the point? Is my product’s next feature going to be redundant in a year, or a month? Can’t AI do what my business does, better?
SaaS valuations have collapsed. From 3-4x revenue to 1x. There’s a sinking realisation that a large number of SaaS companies will go to 0 in the next few years. Building to exit seems almost crazy right now.
There’s far fewer buyers and far more sellers. These include vibe coders cloning products without the care or craft, contributing to the distribution challenges from people doing it the right way.
Lots of people are trying to make money by selling shovels. This just adds to the frenzied energy.
Build in public stopped being fun. A few makers realised that the new game is attention, and now everything feels insincere and stunt-driven. They are influencers not founders. Changes to the X timeline compounded the issues.
There are exceptions and there are still moats remaining, but the challenges feel existential
Yann Lamidon retweeted
We'll be spending a lot more time trying to understand the outputs of language models. A few thoughts, tips & tricks:
Writing. Something I've had success with: Ask your LLM to explain something in ASD-STE100, it's a controlled language specification originally developed for aerospace maintenance documentation. LLMs well-versed in this language and it comes with heavy constraints on clean writing style that I often find a lot more readable. Sometimes I've tried to soften it a bit e.g. ask for "80% of the way to ASD-STE100" because the spec is quite stringent. But even better:
Diagrams / images. Instead of writing, ask your LLM to create a diagram. These can be a lot easier to process, parse, and understand. But even better:
Web pages. Ask for output "in HTML" to get a beautiful, interactive webpage. LLMs are getting really good at frontend and can create beautiful experiences, animations, etc. But even better:
Explainer videos. The output format I am most bullish on is fully custom / bespoke explainer videos generated on any arbitrary topic. Experiment with things like "Create a 3b1b style video explainer on X. Use my ElevenLabs API key for audio narration". (you'd need an API key for the latter or you can ask your LLM to find you decent free alternatives that use your local compute). This is actually starting to work!
In summary:
- As LLMs get better, they will do more and more of the legwork autonomously, and a lot more of our work will rise up the abstractions into oversight and understanding.
- Luckily, LLMs can help here too because as intelligence and code are increasingly abundant, you can ask for large, custom, discardable software artifacts (e.g. web apps, video explainers) that would have never made sense to create before. Push the boundaries here and you'll be surprised.
Yann Lamidon retweeted
The last time building software sped up ~10x (early 2000s, the agile movement), a massive interest in automated testing followed almost immediately (unit tests, TDD, XP etc)
Today, building software sped up 10x, easily: and a similar massive interest in automated testing + verification is following...
Comfy API is live
Point it at a workflow JSON and it handles the rest:
→ Resolves your custom nodes, models, LoRAs, and Python deps
→ Pins it all, ComfyUI version included, into a versioned Build
→ Cuts an immutable release, so what you test is what ships
→ Deploys as an autoscaling endpoint on RTX PRO 6000, H100, H200, or B200
Click the link below to deploy your first workflow and learn more with our blog ⬇️
Yann Lamidon retweeted
We have an official @DeepSeekHarness X account now. Please follow us!
Yann Lamidon retweeted
Introducing Gemini 4 Argon – our new frontier model.
It’s built for complex workflows across coding, enterprise knowledge work, and cybersecurity defense – rolling out today to a set of trusted testers through our Fairwind Program.
Yann Lamidon retweeted
Wow: @petermattis has quite the history in shipping: co-created GIMP, built Gmail’s original backend, shipped Google’s Colossus, and is the cofounder and CTO at Cockroach Labs. We talked distributed DBs:
Timestamps:
00:00 Intro
02:42 Peter’s path into tech
04:00 Building GIMP
09:30 Working on Gmail at Google
14:51 Google’s infra: google3, build files, Bazel, and Colossus
21:30 Distributed storage bottlenecks
23:59 Latency, throughput, and availability
30:04 Contributing to libraries
41:52 Google Spanner
46:10 CockroachDB
52:00 Manual vs. automatic sharding
55:28 Consistency models and strong consistency
1:00:03 Raft consensus
1:06:15 How AI brought Peter back to coding
1:19:12 Peter’s tools and agentic workflows
1:23:08 How AI can improve quality
1:26:39 Code reviews: are they done?
1:29:17 100x engineers
1:35:33 Peter’s advice for leveling up your engineering skills
Brought to you by:
• @turbopuffer – the team are completely redesigning their storage architecture from first principles. And they’re documenting all of it! Follow along,: turbopuffer.com/v3
• @linear – most of us work with agents in a “single-player” setting. Linear’s take is agent work should be teamwork, and they built it so: linear.app/pragmatic
• @WorkOS – fresh from the WorkOS oven: Airlock, the authorization layer for AI agents. It evaluates every request against the agent’s intent and your rules and either allows it, denies it, or routes it to a human for approval. workos.com/airlock
Three fascinating parts from this convo:
1. Peter was recruited to Google because of... GIMP!
The very first version of the famous Google logo was made in GIMP, the free, open-source raster graphics and image editing software created by Peter and his college roommate Spencer Kimball.
In 2001, Sergey reached out to Peter, invited him to an interview and then made an offer. While Peter said no back then (thanks to the commute), he joined a year later, in 2002.
2. Peter twice “beat” standard library data structures in performance terms.
At Google, one of his colleagues noticed that std::map showed up in memory profiles. Peter looked closer and figured out that the std::map implementation is a red-black tree, meaning every node has two pointers. Peter built a B-tree with nearly the same semantics which was both faster, due to spatial locality, and smaller because it used fewer pointers. They used this data structure inside Google.
Peter also did something similar with Go’s map: it was performant, but he built a Swiss Table implementation that was faster. That implementation later made it into the Go library, with the Go team helping to finish it!
3. If you squint hard, everything in distributed databases and storage systems starts looking like a B-tree.
These B-trees are a recurring theme in this podcast episode: the backend of Gmail, the std::map replacement, CockroachDB’s range index, etc. There’s even a paper on this phenomenon, The Ubiquitous B-Tree
Yann Lamidon retweeted
Ideogram 4.5 edits even better than GPT Image 2.5. And it will be open weights! Can't wait to run this locally
I've been actively fighting almost guaranteed age-related loneliness for years, studies show after age 30 you start seeing friends less and when you get older especially men don't have any social life left!
It's preventable, but it takes effort and initiative
We've now organized 70 weekly coworking days where we meet with friends and work together, do show and tell, and try help each other make things (also hardware stuff more recently)
We've also started organizing movie nights, art nights, barbecues, we watched the World Cup final on a projector screen in the garden with friends
We did a collage night the other day where we bought cheap gossip magazines, colored paper, and Pritt glue, and gave everyone scissors to make their own art, I think nobody spent time on their phone for at least 4 hours, making things puts you in a flow state and it's fun to do it in a group because you talk about different stuff while making things than you would for ex having a dinner
Movie nights are nice too but less popular I think cause everyone prefers to watch movies alone and also you can't really talk during a movie
Coworkings range from having sometimes 20 people to sometimes just 2 people, people have their own lives, but me and @rameerez still try to organize them weekly
One thing I realized is that we are usually the ones organizing stuff, but that's not a bad thing, there's even studies about this, there's usually a few people in a group who organize most of the events
If nobody in your group organizes stuff, it doesn't mean nobody wants to do stuff, but you have to take the initiative to bring people together which isn't easy in a time where phones and apps and the internet are simply easier and cheaper dopamine than driving to a friends house for a party
You can start with a coworking day, everyone brings their laptop, and you provide WiFi and coffee, then later you can try more types of events
Having an active social life is just as important or in some studies even more important for health than physical fitness
After a nice party I feel tired but satisfied, I had fun and didn't spend time on my phone or computer but instead talked and laughed with friends etc.
We don't drink alcohol with very few exceptions and nor do our friends, so it's mostly just having really nice food and sparkling water or kombucha etc, and we don't really need alcohol, after a few hours it gets fun enough anyway!
So I recommend everyone, take initiative, and organize something and get people together, it'll be fun! 😊
Great article on increasing solitude of our societies, in this case US but same probably in EU to a lesser extent:
"Derek Thompson's essay "You Are No Longer Invited to Dinner" argues that Americans have largely stopped having people over.
The headline stat comes from combining the old DDB Needham Life Style surveys Robert Putnam used in *Bowling Alone* with a new replication by Data For Progress.
The share of Americans who host friends or family at home at least monthly fell from 42% in 1975 to 12% in 2026, roughly a 70% drop. Most other forms of socializing also fell by double digits over the last 30 years, including cards, bars, sporting events, and volunteering. The decline in hosting was steeper than the much-discussed drop in church attendance.
He rules out two comforting explanations. People haven't just moved dinners to restaurants, since solo dining is rising and face-to-face socializing is down overall. They also haven't replaced parties with other social activities. Exercise and travel are up, but exercise is mostly solitary, and time spent at home has jumped sharply since the 2010s.
His explanations:
1. Harried dual-earner households. As societies get richer, leisure starts to feel like work, and planning a dinner party is basically unpaid project management. Women historically ran the family social calendar. When they moved into paid work, men didn't take over that job. He admits this can't be the whole story, because total leisure time has actually increased.
2. Intensive parenting. Parents, especially fathers, now spend far more time with their kids, and that time comes out of evenings that used to go to hosting adults.
3. Shrinking friend networks. Close friendships are declining, most sharply among people without college degrees, so hosting is becoming a luxury good. People who never host skew toward independents and non-graduates. He speculates that weaker social ties may lead to weaker party attachment, rather than the other way around.
4. Home is too comfortable. Americans spent almost two extra hours a day at home between 2003 and 2022, much of it on screens and short-form video.
His closing idea is a "Lump of Leisure" theory, a twist on the Lump of Labor fallacy. Technology hasn't made human work unnecessary, but it has made other people less necessary for fun. Social leisure turns out to be finite and easy to displace.
He frames this as a coordination failure rather than a personal one: a dinner party requires schedules, cleaning, food, allergies, and the risk of a dull evening, while TikTok asks nothing of you.
A footnote adds that if everyone gets a bit busier, the difficulty of coordinating a get-together rises multiplicatively, so small increases in individual busyness could produce large drops in socializing overall."
derekthompson.org/p/the-deat…
Yann Lamidon retweeted
1/13 Can we assess AI consciousness without first solving consciousness itself?
New from Google DeepMind and collaborators across CS, neuroscience & philosophy:
“From cacophony to hierarchy: a principled framework for assessing AI consciousness”
arxiv.org/abs/2609.35618 🧵
Yann Lamidon retweeted
🔥Welcome 🔺Simplex Diffusion Models🔻!
Standard discrete diffusion models suffer from "information collapse" because they discard uncertainty at intermediate steps by sampling categorical tokens. (1/5)
Yann Lamidon retweeted
It's been almost 2 weeks since we've launched! And due to overwhelmingly popular request, we're making the datasets released with evals.typesafe.ai easier to work with!
We do this in the spirit of openness, but I am still anti-public benchmarks. In that spirit, we deprecate all datasets we evaluate on (internally or publicly). (1/3)
Yann Lamidon retweeted
I updated the AI Engineering Field Guide.
It now includes 6,964 job descriptions analyzed over 8 months (January to August 2026), plus real interview experiences and stories from practitioners.
It covers:
- What AI engineer roles actually look like
- Skills that keep showing up in hiring
- Interview questions and patterns
- Learning paths + resources
It's a work in progress, and I'm adding more sections regularly.
If you have feedback (or want to contribute), I'd love your input.
Repo: github.com/alexeygrigorev/ai…
Star it to get updates!
ALT A detailed AI Engineering Field Guide covering 6,964 job descriptions, skills analysis, responsibilities, and market trends.
ALT A guide outlines learning paths for AI engineering roles, portfolio tips, and job market data from various global locations.
ALT Detailed guide on AI engineering interview preparation, covering processes, question types, skills needed, and trends in hiring.
ALT A guide on AI engineering resources, including interview stories, system design, and webinars on AI careers.
Yann Lamidon retweeted
We turned Qwen3.8-27B into a multimodal decision model.
It beat Pokémon FireRed’s elite four and champion with sub-100 ms decisions from live game state.
With SGLang’s native /v1/decisions, you can now turn LLMs and VLMs into classification and scoring models.
We also added /v1/systemone so Jev-like open models can work with the TypeSafe SDK.
Yann Lamidon retweeted
Took a while, but here it is, a mega write-up on language models for text classification and, of course, Jev.
It's basically a visual guide to RNNs, CNNs, transformers, and calibration, with hands-on experiments on accuracy and efficiency.
Yann Lamidon retweeted
My conversation with Noah Shinn (@noahrshinn), founder of Instinct.
Noah is building a personal AI assistant. It's still invite only, has spent nothing on marketing, and is growing roughly 10% A DAY.
This is his first long conversation about the company.
We discuss:
- Why Instinct doesn't have an app
- Buying compute months ahead of exponential demand
- How users learn to trust it with a credit card
- Safety and security
- Agents coordinating with other people's agents
- Instinct's business model
- Apps built on consumer inertia
- and more
Enjoy!
Timestamps:
0:00 Intro
4:11 What people are using AI agents for
15:07 Rethinking travel, reservations, and the internet
22:43 Trust, privacy, and personal data
27:50 The business model behind Instinct
38:04 How existing businesses will adapt
47:55 Designing a personal assistant people love
53:15 Growth, compute, and competing with Big Tech
1:11:44 What’s next for Instinct and personal AI
Adam Neumann on why top-down management fails with great talent:
"Top to bottom means you think because someone reports in to you, they have to do what you say."
"But it's not true. This is not a dictatorship. They have free will."
"If you hired someone, they're really talented, they'll go work elsewhere. The more talented they are, the more the top to bottom does not work. They won't be willing to do it. They shouldn't."
"Great talent is not willing to be managed like this."
"If you're gonna manage these employees and you wanna attract the best talent in the world, remember that power comes from influence, not control."
"If you think they need to do what you said because you're their boss, you've already lost, and it's a matter of time till it doesn't work out. If they're okay with it... they're not the right employee for you, and you're not the right boss for them. There's no chance you're getting the best out of them."
@AdamNeumann w/ @StevenBartlett
Yann Lamidon retweeted
starting a new week of sharing notable papers, this series topic is latent communication 🤓
batch 1: the pre-history, where models first started passing each other something richer than text
[1] LLM Augmented LLMs: Expanding Capabilities through Composition (CALM) by Bansal et al, 2024
arxiv.org/abs/2401.02412 composes two frozen models with cross-attention on intermediate representations, the direct ancestor of everything below
[2] Let Models Speak Ciphers (CIPHER) by Pham et al, 2023
arxiv.org/abs/2310.06272 multi-agent debate where messages are embedding-space signals derived from output distributions, the first clean argument that tokenizing a message throws away information the other agent could use
[3] Training Large Language Models to Reason in a Continuous Latent Space (Coconut) by Hao et al, 2024
arxiv.org/abs/2412.06769 feeds the last hidden state back in as the next input embedding, so a model reasons without decoding (one model talking to itself)
So excited to welcome @theworldlabs and @drfeifei to the @AMD family! I’ve always been a huge fan of Fei-Fei and her pioneering research in AI. Together, we’ll combine World Labs’ deep expertise in AI and world models with AMD’s compute leadership to power the future of AI and strengthen the open AI ecosystem. Can’t wait for all we’ll accomplish!
