rewatching love island from the start
- Tweets 240
- Following 72
- Followers 13
- Likes 80
sophs h, retweeted
Only 10 days left to enter the Aviation Week Photo Contest. Whether you’re a pro or a passionate spotter, share your most compelling aviation or aerospace image for a chance to win a Tamron 150-500mm f/5-6.7 Di III VC VXD lens ($1,499) + up to $400 cash. Submit to any of our five categories—Commercial Aviation, Defense, Business and General Aviation, Space, and Photo Illustration—and enter as many as you like. Deadline: October 12, 2026.
Enter now: aviationweek.com/photo-conte…
#AviationPhotography #Aerospace #AvGeek #PlaneSpotting #PhotoContest #AviationWeek
6
16
3,746
sophs h, retweeted
Kierkegaard said it best when he said, "Subjectivity is truth."
This is an unfortunately misinterpreted quote. Many think it's arguing for some form of solipsism, but it's not. It's stating that facts are dry and meaningless without a subjectivity, a "soul" if you will, to come along and imbue those facts with meaning. This meaning, we call truth.
Nietzsche would also base his philosophy around this idea.
All an AI can do is mimic meaning. It believes in no truths. A human can come along and completely reprogram (oh, sorry, "re-engineer" or "re-train" are the new terms) the AI's persona to fit some other sense of meaning. Humans, on the other hand, can believe truths that cannot be dislodged short of death: religious martyrs, persecuted scientists, and prophets can all attest to this.
The gulf between humanity and a machine is infinite.
1
1
5
272
sophs h, retweeted
Constrained decoding for diffusion LLMs done right.
Great work by my student @meihuadang : viewing finite automata as graphical models to enable exact constrained decoding for diffusion LLMs, with big improvements on function calling and JevBench.
Very timely given the rise of Jev-like models. To appear at NeurIPS
Introducing Mosaic 🧩, a constrained decoding framework for diffusion language models.
- significantly improves DLMs' performance on general function calling and JevBench.
- compatible with vLLM, SGLang and HF, supporting DiffusionGemma, LLaDA2, and many others.
1/n 🧵
9
11
127
10,187
sophs h, retweeted
Transformer Architecture
Explained in 3 ways here:
- Encoder only
- Deconder only
- Encoder - Decoder both
23
108
4,474
sophs h, retweeted
There are no shortcuts to great work.
We built Google Flow with and for creatives as a tool to use, stretch, remake, and extend their own processes.
And today, we are announcing a series of interactive virtual workshops to spotlight the craft behind modern creative workflows.
Bring your talent. Bring your care. Long live the process.
RSVP below.
24
37
278
216,515
sophs h, retweeted
the number of times i've wanted to tell an agent "debug this codebase, no mistakes"
well, we made a benchmark for that! introducing our newest benchmark, SWE-sweep!
Can LMs discover & fix bugs if you don't tell them what went wrong? To proactively maintain a repo, agents need to find problems before users do.
SWE-sweep benchmarks this on 100 repos, 22 languages, 4k real bugs (numpy, php interpreter, lean kernel). Top models get <5%.
1
4
40
3,173
sophs h, retweeted
🍣 Sushi hybrid quant can now run with MLX-Serve and +50% prefill for M5 Ultra reaching 5,000 tok/s!!
p.s. a known bug for unable to load 🍣 sushi-4bpw ngram table pack, change your metadata group_size from 0 to 1 will work as a workaround.
MLX-Serve 26.10.1 is out: speed across the board.
Exactly faster. Qwen3.8 27B with its drafter
+66% on an M5 Ultra,
+28% on an M4 Max,
+37% on an M1 Pro.
Qwen Flash Next:
+11% decode on M4 Max,
+51% prefill on M5 Ultra. (up to ~5400 PP!!!!)
Gemma 4, Qwen, LFM2.5, Spark, Bonsai 2 all faster, output byte-identical to 26.9.6 (across 18 models).
New:
* GGUF models run on our own MLX engine, written in Zig and Metal, instead of the llama.cpp fallback. (experimental)
* 🍣 Sushi-format Flash Next packs run directly.
* Qwen-Image 2.1 edits pictures from instructions.
* Drafting is on by default for every model that supports it.
Huge thanks to everyone who filed, tested and fixed this one: @STRML_, @sbusso, @wutang_superfan , lborloz, LXD-8 @alinselea zeeshanhaque21 @Lojza3D, h9q2cyxvgm-ui @kennethrdegraff, @CowboyCoderHQ , codysk, @Beamsters1, @CerebralCoding_, @AjAbsaki , and more.. THANKS ! Please comment if I did not include your name, for visibility, I know people mostly by their GH handle.
Special thanks to @ashxhart and @Spangler3000 for pushing !
Benchmarked and tested across 5 machines: htmlhost.jax.workers.dev/ren…
GH Release Download + Changelog:
github.com/ddalcu/mlx-serve/…
Website: mlxserve.com
10
4
39
4,112
sophs h, retweeted
An annoying feature of @ChatGPT for Clinicians is its refusal to do any arithmetic related to a medical question, not present in the free consumer version.
5
3
24
5,546
sophs h, retweeted
Ceiling of a Task: When Can a Transformer Succeed Without Its Chain of Thought?
The authors show that a transformer's chain of thought is computationally necessary for serial tasks; removing it drops accuracy down to a shallow-circuit ceiling.
arxivb.org/abs/2609.33134
3
3
45
2,411
sophs h, retweeted
i think this is the biggest advantage of living in Bay Area 🇺🇸; when I discuss here the economics of post-AGI companies people don't even blink; during my euro-summer 🇪🇺 (nearly) everyone just laughed cause they thought I'm joking
Partly. Monocultures are also productivity machines.
When everyone around you already accepts the premises, you stop burning energy defending them and can push an idea much further than you could anywhere else.
Plenty of extreme bets only get made when nobody around you makes you question the premises.
Doesn't make the post wrong: You want one foot inside the monoculture for the leverage, and one foot outside so it doesn't decide what you want.
1
4
38
3,087
sophs h, retweeted
Les gars, ça fait 5 jours que je suis à San Francisco. J’ai parlé à des dizaines de personnes, notamment beaucoup de Français dans la baie, dont beaucoup sont revenus ou sont arrivés récemment.
Ils font tous le même constat, avec différentes nuances et différents raisonnements : la France, c’est mort. L’Europe a échoué.
Une crise sans précédent arrive.
Il n’y a plus de solution. Il faut qu’on touche le fond pour tout défoncer et reconstruire de zéro. Et accepter l’échec.
On ne se rend pas compte de l’enjeu de l’IA , de l’inférence , du besoin d’avoir des datacenters.
La prochaine élection présidentielle est une immense blague, avec des candidats qu’on se mange depuis déjà 20 ans et qui nous expliquent ce qu’il faudrait faire, alors que beaucoup ont déjà été au pouvoir.
Les élites s’enrichissent entre elles, et les gens qui parlent de tech et d’IA n’ont absolument aucune idée de ce dont ils parlent.
Les politiques sont totalement largués sur les sujets d’IA, et on a déjà loupé 15 trains.
L’opinion publique raconte et croit n’importe quoi sur l’IA, portée par des médias qui racontent eux aussi n’importe quoi.
Des personnes totalement incompétentes, qui ne savent pas de quoi elles parlent, relaient en continu de la peur autour des nouvelles technologies, et ça contribue au retard qu’on est en train d’accumuler.
En vrai, ça me fait beaucoup cogiter d’être ici. Bref je continue mon petit tour.
128
153
1,283
109,425
sophs h, retweeted
Ready for the final day? 🙌
There’s still time to discover the TASKalfa Pro 15000b and TASKalfa Pro Tandem at Booth N5043!
📍 @PRINTINGUnited
#PRINTINGUnitedExpo #PRINTINGUnited #PRINTINGUnited2026
2
1
82
sophs h, retweeted
What an absolute banger!
@dhh is an artist on delivering ideas in exciting and fun way.
If haven't watched @dhh rails world talk you should take an hour to see it.
It's pencils down, people. Writing code by hand is no longer an economically viable skill for most programmers at most companies. But the future of making software has never been brighter. Don't you dare black pill this beautiful moment! piped.video/vDjW_dRyKXY?si=6Fsf…
1
2
285
sophs h, retweeted
Your hemorrhagic shock pt got pre-hospital blood, but does your system have the QA and blood bank infrastructure to back that up?
Mike Carunchio, FP-C (@okmedicpodcast) from ResusX:2026:
The product is only as good as the program behind it. How does your shop handle this?👇
2
2
777