Graham Fleming

3.8K posts

Graham Fleming banner
Graham Fleming

Graham Fleming

@GrahamFleming

Mechatronics Engineer | Human Advisor @NeuralHub_infra | @vercel

Katılım Aralık 2013
1.9K Takip Edilen3.8K Takipçiler
Sabitlenmiş Tweet
Graham Fleming
Graham Fleming@GrahamFleming·
Physical AI Hack 2026 at Founders Inc in San Francisco Builders came together for a robot hackathon with real hardware like the LeRobot SO-100, LeKiwi, Unitree G1 and others. Teams tackled live manipulation challenges picking from puzzle and shape insertion, plugging in chargers, and pouring liquids into cups. We explored transfer learning, fine tuning VLMs and VLAs, closed loop policies, and generalization across messy environments all with visible, measurable progress that you can't fake. Physical AI is bridging the gap from digital models to tangible impact faster than ever, but the real magic happens when hardware meets adaptive software. Shout out to organizers @fdotinc, Dhruv Diddi, Devinder Sodhi + more
Graham Fleming tweet mediaGraham Fleming tweet mediaGraham Fleming tweet media
English
3
1
17
1.4K
Dave W Plummer
Dave W Plummer@davepl1968·
Hard Drive activity indicator for the Mac: my first Swift app. Wrote it this morning because I wanted it. It monitors disk queues at 120Hz and triggers the LED for 10ms, which seems pretty realistic. What should I do with it? I pay the $100 Apple Developer tax, should I put it on the app store for $1 or just stick it on my GitHub and hope people can figure it out?
English
161
30
1.3K
238K
Graham Fleming
Graham Fleming@GrahamFleming·
@fchollet Used to be only deterministic predefined workflows that did one thing. Now software is becoming more general and dynamic, molding to the users needs to take action.
English
1
0
0
81
François Chollet
François Chollet@fchollet·
Thinking of AI as a productivity booster for prior workflows is the wrong framing. Like all of the previous waves of computerization/softwarization, AI is a tool that lets you do new things in new ways.
Computers and Society Papers@WGOV

Cognitive offloading and the speedup illusion in human-AI interaction Sunny Yu, Myra Cheng, Ahmad Jabbar, Ilia Sucholutsky, Katherine M. Collins, Dan Jurafsky, Robert D. Hawkins arxiv.org/abs/2605.23177 [𝚌𝚜.𝙲𝚈 𝚌𝚜.𝙷𝙲]

English
60
78
578
58.3K
Graham Fleming
Graham Fleming@GrahamFleming·
AI injected "zero width characters" into PRs and cooked
Socket@SocketSecurity

🚨 BREAKING: Active supply chain attack across npm, PyPI, and Crates.​io. Socket detected TrapDoor, a crypto stealer campaign hitting 34 malicious packages and 384 versions and artifacts, with attackers repeatedly pushing new releases across ecosystems. TrapDoor targets #crypto, #DeFi, AI, and security developers, stealing wallets, SSH keys, cloud credentials, GitHub tokens, browser data, env vars, and API keys. Socket detected releases with a median detection time of 5 minutes, 27 seconds. The fastest detection occurred 58 seconds after publication.

English
0
0
2
126
Graham Fleming
Graham Fleming@GrahamFleming·
Iterative analysis and engineering of itself
Qwen@Alibaba_Qwen

Self-Evolving in the Wild：Over the course of ~35 hours of continuous autonomous execution, the model performed 432 kernel evaluations across 1,158 tool calls. It wrote, compiled, profiled, and iteratively improved the Extend Attention Kernel entirely on its own — 10.0x geometric mean speedup over the Triton reference, measured across multiple workloads. More details on blog:qwen.ai/blog?id=qwen3.7

English
0
0
2
145
Graham Fleming
Graham Fleming@GrahamFleming·
@fchollet Lots of AI will exist just to generate data for other AI No UI needed
English
0
0
0
25
Graham Fleming
Graham Fleming@GrahamFleming·
Looking at different radio communications for the cubesat project. UHF 435 MHz is the standard with the longest range at 1000 km and reasonable data speeds and low power. LoRa 2.4 GHz is newer and can send more data, but with shorter ranges and higher power Will keep digging
Graham Fleming tweet media
English
1
0
2
67
Graham Fleming
Graham Fleming@GrahamFleming·
@fchollet This can be done by giving more context and inference Although whoever can do it the quickest and most efficiently wins
English
0
0
0
128
François Chollet
François Chollet@fchollet·
Most human tasks are not Markovian, the optimal next action cannot be determined solely by looking at the current state. It depends heavily on the past trajectory, the original intent, and context constraints. An agent that cannot compress and track its past trajectory with absolute fidelity is maybe 20% as useful as one that can.
English
90
70
963
54.6K
Ivan Burazin
Ivan Burazin@ivanburazin·
2.5% of our sandboxes run longer than 24 hrs. That 2.5% brings 20% of our revenue. Long-running stateful workloads are not an edge case. It feels weird to see that this isn't the consensus yet.
English
17
3
117
10.9K
Ivan Burazin
Ivan Burazin@ivanburazin·
I was terrified of public speaking. Used to go to the washroom 10 times before any stage appearance. Then, a few years ago, we threw a 250-person conference, and the MC bailed at the last minute. I had to MC two full days in a row without a choice. Otherwise, the whole thing would have collapsed. Being forced into it with no exit broke my genuinely crippling stage fright. I'm p sure no amount of practice or courses would have helped as much.
English
13
2
78
5.4K
Graham Fleming
Graham Fleming@GrahamFleming·
the AI model wars are becoming psychological
Udi Wertheimer@udiWertheimer

claude just MOGGED me for using codex over the weekend on saturday morning i was brainstorming with claude about a weekend project. made some interesting observations and found some stuff worth exploring then i switched to codex and had it work through research and prototyping most of the weekend, and ended with a thorough implementation plan for a large project last night i came back to claude, showed it the plan and the other artifacts, said “hey check this out i worked on this with codex all weekend what do you think of the plan?” and damn i’m sorry i asked “you didn’t plan any of the hard stuff. you polished the lobby so it looks good to visitors but didn’t architect any other floor” “the tell is that the thoroughness clusters exactly where the thinking is enjoyable and thins out exactly where it’s grim” “the product is strictly worse than the status quo” “you push back hard on hollow hype when other people do it; your own doc is doing the same thing you hate” “this is exactly why you shouldn’t have spent the weekend on this” “if you want a concrete next step: throw all of this away” BRO WOW

English
0
0
0
81
François Chollet
François Chollet@fchollet·
A mental model for working with coding agents is that they're blind squirrels running into a maze and bumping into walls. You must place the walls (verifiable constraints) strategically so that they end up in the general region you want them in.
English
106
161
1.5K
123.9K
Nikita Bier
Nikita Bier@nikitabier·
Pro-tip to video creators: Add captions to your videos. Most people scroll Timeline with volume off. It will increase engagement on your content by 5x. We’ll be adding this natively to our video uploader soon.
English
2.3K
1.1K
14.3K
907.5K

Keşfet

@davepl1968 @nikitabier @fchollet @justalexoki @ivanburazin @elonmusk @BarackObama @taylorswift13