Igor Susmelj (@ISusmelj) header
Igor Susmelj (@ISusmelj) profile picture

Co-founder @lightlyAI | Speaker | @ETH | Blogging about ML and data

Zurich
Joined March 2019
Photo by Igor Susmelj (@ISusmelj)Photo by Igor Susmelj (@ISusmelj)Photo by Igor Susmelj (@ISusmelj)Photo by Igor Susmelj (@ISusmelj)Photo by Igor Susmelj (@ISusmelj)Photo by Igor Susmelj (@ISusmelj)Photo by Igor Susmelj (@ISusmelj)Photo by Igor Susmelj (@ISusmelj)Photo by Igor Susmelj (@ISusmelj)Photo by Igor Susmelj (@ISusmelj)
3D point clouds are coming to LightlyStudio πŸ‘€ MCAP support, different projections, point cloud accumulation, object tracking across frames, and more. Still experimental, but getting close! Working with 3D data? We'd love your feedback. Sneak peek πŸ‘‡
1
1
71
Model evaluation in LightlyStudio just got easier πŸš€ Run + rerun evaluations directly in the UI. Click a confusion matrix cell to inspect failures, or sort by FP/FN to find the worst cases. No switching between notebooks and visualization tools. eu1.hubs.ly/H0zBJyr0
1
1
87
LightlySSL v1.6 is out! πŸ€– We're expanding our OSS framework from self-supervised vision learning to world models. First up: LeWM. Predict future visual embeddings conditioned on actions, then use them for planning. Includes a hands-on tutorial! eu1.hubs.ly/H0zpWhg0
1
1
136
What if you could steer your embedding visualization with natural language? Coming to LightlyStudio: text-guided axes for 2D embedding projections. Explore your image and video datasets along concepts you actually care about. Next release πŸš€
23
26
437
My daily routine now: Wake up β†’ check agents β†’ breakfast β†’ check agents β†’ office β†’ check agents β†’ home β†’ check agents β†’ gym β†’ check agents between sets πŸ˜… Claude + Codex are great at scoped tasks, especially code optimization. But manager agents? Often a flood of slop that's harder to clean up than doing the work yourself. Feels like we've automated coding, but not yet managing the coding agents.
1
77
We've been improving model evaluation in LightlyStudio πŸš€ Compare predictions vs. ground truth, inspect errors, adjust IoU thresholds, and rerun evaluations. All directly in the UI. Much easier to understand where your model fails. Quick demo πŸ‘‡
1
141
AI-assisted labeling is coming to LightlyStudio πŸš€ First up: superpixel-based segmentation to make drawing masks much faster. The cool part? Everything runs directly in your browser. No GPU servers, APIs, or complex backend. More labeling features coming over the next few weeks πŸ‘€
82
Been playing around with EPIC-KITCHENS-100 in LightlyStudio. Indexed 37K+ egocentric video clips and explored them using embeddings. Pretty fascinating how quickly you can discover patterns, near-duplicates, and mismatches between narrations and actual video content. Tutorial πŸ‘‡eu1.hubs.ly/H0yZ_Wc0
3
95
New LightlyStudio tutorial πŸš€ Explore ~10K images using embeddings, select clusters with the lasso tool, review + bulk label them, and evaluate against ground truth. Try it yourself πŸ‘‡ docs.lightly.ai/studio/tutor…
1
49
Pretty fascinating to watch agents optimize other agents. Gave GPT-6 Astra the system prompt + tool descriptions for a cheaper flash model, an eval set + expected outcomes, and told it to iterate twice while keeping costs low. 2/7 β†’ 5/7 β†’ 6/7 satisfactory answers. Cost stayed basically flat. Then I just let it work.
89
4.5M images in LightlyStudio πŸ‘€ Embeddings, interactive visualization, filtering and sampling, all on the same dataset. Rust really paid off here, especially for performance + memory usage. We’ve done 1M+ videos before. Next target: 10M+ samples. Let’s see where it breaks πŸ˜…
65
We just added bulk labeling with embeddings to @LightlyAIStudio πŸš€ Open your data β†’ embedding view β†’ lasso a cluster β†’ label the whole group. Works for object embeddings too. Great for splitting classes (person β†’ adult / kid) and annotation QA, where wrong labels often pop out as outliers. Demo πŸ‘‡
61
Doing some larger pre-training runs again at @LightlyAI. Pretty amazing how much the training stack has improved. WebDataset, activation checkpointing, reduced precision, 8-bit optimizers, torch.compile... things that used to be painful are now surprisingly robust. Stay tuned πŸ€–πŸ¦–
4
123
I’ve been a bit silent here, but definitely not idle πŸ˜… We’ve been onboarding customers and shipping a lot at @LightlyAI. One thing I never got around to sharing: our revamped LightlyStudio docs. Pretty proud of how far this has come πŸš€
1
4
475
πŸ’Έ Easy way to save money on GPT / Claude / Codex usage: stop using the same chat forever. Old context = more tokens, more cost, more noise. My rule: new day = new session, new topic = new session. For bigger projects, keep the state in a context markdown file. I used 50M+ tokens in 7 days and still had ~50% of my weekly quota left.
166
The price of intelligence keeps collapsing. DeepSeek V4 Flash is now nearly 20x cheaper than Gemini Flash 3.6 and around 100x cheaper than leading OpenAI and Anthropic models. Soon, using an LLM to classify text may be cheaper than training a classifier.
1
4
184
When new backbones drop, everyone asks the same question: how well do they transfer? We compared three on ADE20K segmentation using LightlyTrain: 🟒 LingBot: 0.465 mIoU πŸ”΅ EUPE: 0.456 ⚫ DINOv3: 0.449 LingBot comes out on top πŸš€ And yes, all three are supported in LightlyTrain.
1
2
354
Open source doesn’t just release models. It distributes the optimization problem to thousands of smart people. Kimi K3 is already being pushed onto Apple Silicon and lower-specced hardware. Interesting project I saw on Hacker News today: eu1.hubs.ly/H0xh89L0
1
132