jasperanders (@_asperanders) header
jasperanders (@_asperanders) profile picture
Joined April 2021
Photo by jasperanders (@_asperanders)Photo by jasperanders (@_asperanders)
Did you ever see someone flipping the sign of your loss, and achieve great results? Well, this summer, we did. Some papers (refs below) suggested to reverse the self-distillation loss, i.e, move away from the teacher. We wrote a blogpost on our investigation and our findings.
Self-distillation with privileged context shifts behavior and can hurt reasoning. Repulsion reverses the shift but becomes unstable. In our experiments, contrasting both signals improves reasoning while keeping training stable. Blog: self-distillation.github.io/…
1
4
14
1,231
We needed VR teleoperation for intervention-based learning. So Aurel implemented the entire stack within two days. He now wrote an extensive blog post on the design decisions and embedded math for Pieper IK. We also open-sourced the code. Highly recommend for anyone into IK.
I implemented VR teleoperation for our robot arm because I wanted to do human intervention data collection and our GELLO leader arms didn’t allow it. I wrote all my learnings up and open-sourced the code…
4
3
26
3,845
jasperanders retweeted
es tut mir so leid, dass auf den Essigessenz-Flaschen immer Gemüse drauf ist…Bro…du kommst in keinen Salat…du entkalkst meinen Wasserkocher…
72
142
16
5,961
118,366
me reviewing codex’s output after it worked for 16 hours straight
13
18
2
389
30,959
jasperanders retweeted
Today we’re releasing DeepSWE, a new standard for agentic coding benchmarks. On public leaderboards, top models often look relatively close in capability. DeepSWE shows where they actually diverge, reflecting the realistic experience of developers in their day-to-day work.
515
718
419
6,013
1,989,049
Makes so much sense to share things this way in a AI native era.
I built .bdf to easily share html directly from my agents. I use Claude a lot to create html slides and documents, but sharing plain html with others is annoying. .bdf gives you a shareable link in a single prompt. give it a try: bdf.pub
1
16
4,009
Sharing our latest work: Reinforcement Learning via Self-Distillation We introduce Self-Distillation Policy Optimization (SDPO), an on-policy RL algorithm that turns feedback into dense, token-level learning signals.
4
7
4
58
12,219
I've got 2 unwritten rules. 1. 2.
42
152
11
1,153
jasperanders retweeted
irgendwann hast du zum ersten Mal eine E-Mail empfangen und es hat seitdem nicht mehr aufgehört, es wird nie mehr aufhören, irgendwann hört dein Herz auf zu schlagen und die E-Mails werden weiterhin kommen
16
65
6
3,261
jasperanders retweeted
3/5 „Es ist jetzt sinnvoll die Menschen vor den Kauf von neuen Benzin und Diesel Autos zu schützen!“
4
5
3
78
jasperanders retweeted
I love finding New Yorker cartoons so dated that the joke is lost entirely and the cartoons become just descriptions of people doing normal things.
102
9,943
255
76,482
Noch mal zur Klärung: Die BVG ist nicht das Bundesverfassungsgericht. Wir halten Notbremsen aber ebenfalls für richtig.
88
953
102
15,307
DER @Tech4Germany JAHRGANG 2021 💙🥳🚀 @HBraun @BAMResearch @BMFSFJ @BMVI @Bundestag @ITZBund @lehnerpat @katja_anokhina @Lergin_ @GroteSophia @NilsThamm @sinado @BKorkmaz_KD @jasperanders3
6
1
29
FELLOWSHIP 2021 🚀🥳💙 Die 7 Projektteams für das diesjährige Fellowship stehen fest! 🤝 28 Fellows & 28 Verwaltungsmitarbeitende werden in nur 3 Monaten nutzer:innenzentrierte #Software-#Prototypen für Bundesbehörden entwickeln. Alle Infos hier 👇 🔗: tech.4germany.org/fellowship…
1
7
2
16