James Assiene retweeted
Stop forcing a 2,700-word prompt + 900ms frontier model to do a fine-tuned smaller modelâs job. đȘ
@AirbnbEng just fine-tuned @GoogleGemma 4 E2B on a single GPU with @vllm_project, to serve as a JEV-like classifier. The results are pretty solid:
⥠35ms median latency (85ms p99)
đŻ 0.902 F1 (matches frontier models)
đ ~1% of the cost at scale
đ Read the full deep dive on GLOWC + their @GoogleDeepMind Gemma 4 setup here:
airbnb.tech/ai-ml/frontier-qâŠ
Just wanted to share a (very much WIP) little side project of mine showcasing the benefits/speed of @googlegemma DiffusionGemma đ (text diffusion model)
Project is called âDiffusionGearvisâ (mix between DiffusionGemma and Jarvis, IronManâs AI assistant) and I wanted to implement some features for which I believe text diffusion can really shine, the first one being computer use.
With the success of Muse from Meta, it is very clear that computer use has tremendous value for users. However, this feature is 1) notoriously slow 2) raises serious privacy concerns (screenshots and other personal data uploaded to the cloud). DiffusionGemma can help with both đ It is super fast thanks to text diffusion and is open-weight so people can run it privately on their own desktop / VM. All the demo videos were run locally on a desktop with a single @NVIDIAAIDev RTX 5090 with speech-to-text and text-to-speech from @kyutai_labs (stt-1b-en_fr and pocket_tts) .
This is with the vanilla NVFP4 checkpoint served on @vllm_project so performance can vastly be improved with dedicated fine-tuning or distillation from larger open-weights models. I hope that in a few months, the open-source community will create some fast âlocal Museâ on top of it :
Everything is open-source so feel free to use/fork/whatever đ github.com/james-assiene/DifâŠ
Another use-case close to my heart is education: it has been proven that thinking significantly improves quality (+ ~20p in accuracy on AIME 2026 according to DiffusionGemmaâs tech report) but with auto-regressive models, it often comes at the cost of latency (more tokens, more time you wait). Thanks to the speed of text diffusion, you can benefit from the quality of âvery thoughtful answerâ without the latency penalty, making smooth and smart real-time assistant use-cases like live tutoring a reality
Lastly, because Iâm a bit of a nerd, gaming ^_^. I think weâve all been in a situation where we are a beginner at a game, or forgot the commands to do XYZ, or needed to go on the internet to unblock ourselves in a game. AI gaming companion are here to help without us switching to our phones/browsers but they can sometimes be slow if the question requires heavy thinking or tool calling (therefore making the experience less smooth). Once again, text-diffusion is here to help. In this case, the game is running on the cloud (@XBOXGamePass cloud gaming) so that the AI models can use all the VRAM.
James Assiene retweeted
Africaâs digital future is now. đ
Weâre launching Skilling Africa with Google to bring 15+ Google training initiatives into one offline-ready learning hub.
Our goal: Train 3M Africans in AI & digital skills by 2030.
Start learning for free today âĄïž skillingafrica.withgoogle.coâŠ
James Assiene retweeted
We asked GPT-6 Astra to edit Qwen3.5-0.8B weights to improve math.
It directly changed only 0.016% of weights, fixing 0.8B's reasoning loops & improving math up to 6x and even instruction following 3x, without gradients or training.
We call this Weight-Space Programming: frontier LLMs can now natively read/write neural network weights! đ§”
Blog: rikiyat.github.io/blog/weighâŠ
Readers added context they thought people might want to know
LLM was given weights explicitly as text in prompt and used code/forward passes (per authors' logs) to ablate Qwen looping direction, a known thinking-mode issue per model card. Primarily fixed repetition in that mode, not general 'natively' improving weights/math.
rikiyat.github.io/blog/weight-spâŠ
github.com/RikiyaT/lm_undâŠ
huggingface.co/Qwen/Qwen3.5-0âŠ
huggingface.co/RikiyaT/Qwen3.âŠ
James Assiene retweeted
The numbers they donât want you to know about immigrant care workers.
1/6
The anti-immigration circus is on a full-on attack, citing numbers they don't understand, and almost all the figures they cite are wrong.
First, the cost.
An 85-year-old woman who canât wash or dress herself receives 14 hours of care a week. This costs the council ÂŁ351 (ÂŁ25.05 an hour). If no care worker comes and she ends up in hospital, the NHS spends ÂŁ3,934 a week (ÂŁ562 a day), which is 11 times the cost.
As of March 2026, an estimated 150,000 direct care posts in Englandâs private and charity-run care providers were filled by Health and Care visa holders. Taking the average working hours in the sector, this comes to about 225 million hours of care a year, worth ÂŁ5.6bn at the rates councils pay.
For every ÂŁ1 spent on social care, the NHS saves between 20p and 35p. So these workers save the NHS between ÂŁ1.1bn and ÂŁ2bn a year.
The total lifetime cost of these care workers and their adult dependants comes to ÂŁ9.5bn. The NHS savings alone pay this back within 9 years.
And if no care worker comes, the family steps in. In March 2026, 1.4 million people were claiming Carerâs Allowance, 56,800 more than in February 2025. It pays ÂŁ86.45 a week for at least 35 hours of care a week. In the UK, 544,000 working-age carers are out of work because of caring. This costs the economy ÂŁ37bn a year in lost output. And this figure only includes those caring for people of working age. A daughter who has left her job to care for her 85-year-old mother isnât even included in it.
So, how much does an unpaid family carer cost the taxpayer? According to UK data from 2017 to 2020, a working-age family carer who provides 10 or more hours of care a week costs the state ÂŁ5,339 a year more than someone in a similar situation who doesnât provide care: ÂŁ3,691 in benefits, ÂŁ1,620 in lost tax and ÂŁ28 in extra NHS costs.
And demand will only keep rising; in England alone, 410,000 more care posts will be needed by 2040, just to keep pace with the growth in the population aged 65 and over. The number of people with dementia in the UK is projected to rise from 982,000 in 2024 to 1.4 million by 2040. Meanwhile, British workers in care are down 130,000 since 2020/21.
James Assiene retweeted
To be a celebrated African INTELLECTUAL, do you:
-develop a new scientific theory. Nope.
- invent groundbreaking technology. Nope.
- whine about colonialism with eloquent English. YES.
We have turned the African voice into a perpetual whine fest. This is why no one respects us around the world. Everyone else got occupied, colonised, invaded, pillaged. Often worse than what Africa went through, and even longer. Those countries have surpassed us, some even surpassing their former colonisers in GDP per capita. Still, all we do is whine.
Can we please stop?
They invited her to Berlin thinking she would praise them. Instead, she exposed their blood-soaked history.
Chimamanda Ngozi Adichie didn't hold back. She exposed the ultimate Western lie. They teach their kids to "help the poor Africans," but completely erase the fact that Belgiumâs King Leopold turned the Congo into a personal slaughterhouse. Millions murdered. Hands chopped off for rubber.
They don't teach the truth because the truth is a debt they can never repay.
Watch this. Bookmark it. Teach your children who the real savages were.
James Assiene retweeted
Incredible "hello world" for Project Suncatcher â and an exciting milestone for the Gemma family we built for embedded / edge workloads!
Three things that make this engineering wild:
1. True edge / embedded inference: Gemma 4 is running completely offline in low Earth orbit â zero ground relay for compute, executing prompts and payload telemetry checks directly on the spacecraft.
2. Data-center silicon in vacuum: Running on 4 radiation-tested Trillium TPUs powered by direct solar energy, surviving launch vibration and thermal extremes.
3. From laptops to orbit: We designed Gemma from first principles to maximize capability per watt in constrained environments â seeing that same architecture scale from local devices to an orbiting satellite is surreal.
Huge kudos to the Google Research, Planet, TPU, and Gemma teams!
James Assiene retweeted
This incredible. Gemma 4 became the first LLM to run in space?
James Assiene retweeted
Nobody ever seems to demand Arab nations to apologise or give reparations for their slave trade, which was even larger, longer, and brutal than the transatlantic slave trade.
James Assiene retweeted
Slavery was an evil in which people from every continent and region participated. We should be honest about the whole history, not select the parts that are politically convenient today, especially given Britain's leading role in ending the transatlantic slave trade.
If Caribbean countries are asking Britain for a formal apology for Britainâs role in the transatlantic slave trade, will they also be asking the same of the descendants of African kingdoms, rulers and merchants who captured, sold and profited from the trade?
On demands for reparations, whether through cash or debt relief, we have been down this track before. Twenty-five years ago, Jubilee 2000, Drop the Debt and Make Poverty History argued that cancelling the debts of poorer countries would give them a fresh start. More than $100 billion of debt was eventually written off, but many of those countries have since built up even bigger debts. Labour are already spending too much on Britain's welfare state, now is not the time for Andy Burnham to be providing global welfare.
There is no reason British taxpayers today should be held financially responsible for things that happened generations before any of us were born. And I have even less patience for political elites in commonwealth countries who donât know how to make their own citizens prosperous now looking for another opportunity to blame Britain.
As Prime Minister, Andy Burnham should be Britainâs greatest defender. Instead, he is writing articles saying we need to confront our history (as if we haven't done enough of that) and considering another formal historical apology. Tony Blair already expressed Britainâs deep sorrow and regret for our role in the slave trade nearly twenty years ago. There is no need to do so again.
This is all another navel-gazing distraction. The Commonwealth should be looking forward, not endlessly backwards. That means economic partnerships to make us all richer and deeper military and security cooperation against common threats. If every Commonwealth summit becomes an exercise in presenting Britain with demands for cash then we need to start asking if it is worth the effort.
James Assiene retweeted
BREAKING: Google just proved AI data centers in space actually work
Project Suncatcher's first satellite, built with @planet and launched on a SpaceX rocket with 4 TPUs on board, verified the physics: Google sent a prompt from Earth, a TPU in orbit ran Gemma inference, and the full answer was beamed back down.
SpaceX is launching its StarMind satellites with NVIDIA GPUs on Starship in late 2027.
Kardashev Type II civilization INCOMING
James Assiene retweeted
@GuillaumeLample how is the RL run coming along?
@GuillaumeLample :
James Assiene retweeted
iâve found myself using Genie 4 to have conversations with younger versions of myself...
hereâs me talking to myself just out of college. i was pretty naive, but not a whole lot has changed :)
James Assiene retweeted
I'm so glad AI killed LeetCode-style coding interviews.
I've talked to young AI talent, and most would simply skip companies that still use LeetCode interviews.
I agree with Felix. Taste, empathy, principled thinking, business sense, and agency are what really matter.
The coding interview is dead, at least for my team at Anthropic. I've seen enough to know that it no longer provides signal. I removed it from loops this month.
I have a spicy, optimistic take - I think there's an opportunity for junior engineers: felixrieseberg.com/hunger-beâŠ
James Assiene retweeted
Speeding up Gemma!
Gemini 4 Argon enabled a 2.85x speedup on an A100 via vLLM, jumping from ~37,900 to 108,000+ tokens/sec on gemma-4-E2B-it.
â
100% token parity
â
Zero quality drop
â
Done in days, not months
See the exact changes across 8 PRs: goo.gle/4zNtANl
James Assiene retweeted
Le racisme Ă lâĂ©gard des Noirs existe.
Le racisme Ă lâĂ©gard des Blancs existe.
Le racisme Ă lâĂ©gard des MaghrĂ©bins existe.
LâantisĂ©mitisme existe.
ArrĂȘtez de penser quâil nây a que votre communautĂ© qui est la victime, car câest exactement ce que les marionnettistes veulent : LA DIVISION.
James Assiene retweeted
"how to avoid brain rot" was a great watch. some of my favorites points:
- using ai to ask yourself socratic questions instead of giving you the answer makes you feel like you're learning less but research shows you're actually learning more
- ai tooling is moving so fast that it's better to know 50% of a new tool than to master 90% of an older tool
