Dejan Kostic

Computer Science Professor ( and a photographer ( Also

[Dejan Kostic @dmkostic]

Deeply honored to become a Wallenberg Scholar! @KAWstiftelsen

https://t.co/3ZJDJAVgEA

[Dejan Kostic @dmkostic]

Very happy to announce that Giacomo Verardo defended his doctoral thesis! @marchiesa was the co-advisor and @GQMaguireJr provided stellar insights. More details and the link to Giacomo's thesis:

https://t.co/SH9OnyRDiU

[Dejan Kostic @dmkostic]

New postdoc opening in my DAI project on creating a highly energy-efficient and cheap platform for AI inferencing! The project blog gives more information on our recent work and the overall project vision:

https://t.co/wRXSPBWZZT

https://t.co/b1uzX2DGz7

[Dejan Kostic @dmkostic]

Spring has arrived at the main KTH campus! #stockholm #sweden #kth

[Dejan Kostic @dmkostic]

New ad for a PhD position, new aurora picture! Large Language Models keep getting bigger and more capable, we keep trying to make inferencing much more efficient:

https://t.co/GN5rSppcMt

[Dejan Kostic @dmkostic]

More details about work, including links to the paper and source code at the blog site below:

https://t.co/1UUjOD1pCb

Xiangyu Shi @GQMaguireJr @marchiesa (5/5)

[Dejan Kostic @dmkostic]

Which form of communication should Large Language Models (LLMs) use when collaborating on a problem, e.g., one or two models processing the context and another model answering a question without seeing the context itself: plain text, hidden state, or some other way? (1/5)

[Dejan Kostic @dmkostic]

Interestingly enough, on two data sets KVComm even exceeds the performance of an approach that gives both the context and the question in text to the model answering the question. Our approach works even for different LLMs that are fine-tuned from the same base model. (4/5)

[Dejan Kostic @dmkostic]

Happy to announce that Daniel Perez successfully defended his PhD thesis on February 12, 2026! This thesis uses ML to solve resource allocation optimization problems and demonstrates 100x scaling in problem size. https://t.co/qNLasEBR1N

[Dejan Kostic @dmkostic]

Absolutely magical on the main KTH campus these days (image from Sunday)… #kth #Stockholm #Sweden

[Dejan Kostic @dmkostic]

Beautiful October evening in Stockholm, Aurora and reflections…cannot ask for more!

[Marco Chiesa @marchiesa]

Congratulations @WangChangjie13 on receiving the 2025 Applied Networking research prize for our NetConfEval work! 🏆

Now, let’s work on making LLMs also reliable and sustainable for code generation! 🚀

https://t.co/rh55z4F67V

[Dejan Kostic @dmkostic]

3/3 This happens because possible gains get dwarfed by slowdowns in accessing memory.
@ghasemirahni @alirezafarshin @marchiesa @GQMaguireJr

[Dejan Kostic @dmkostic]

1/3
At CoNEXT ’24, @Skazza_94 presented our FAJITA paper for running stateful network functions at very high rates (~1.4 Tbps). PDF + video are at the @ERC_Research CoG ULTRA blog post below. Click below to see something perhaps unexpected in this work.
https://t.co/P0B47cj4tu

[Dejan Kostic @dmkostic]

2/3
Unless the number of so-called “elephant flows” is very small, spreading incoming packets among the cores using plain Receive Side Scaling (RSS) outperforms existing approaches that perform fine-grained flow accounting and load-balancing.

[Dejan Kostic @dmkostic]

@axiomaticbin Hi, the link is in the first post of the thread (and it’s free to apply)

[Dejan Kostic @dmkostic]

I am continuing to look for two more PhD students who wish to work with us on creating a dramatically more energy efficient and cheaper platform for AI inferencing (deadline Jan 17, 2025):

https://t.co/Gb9KXk8xhY

1/3

[Dejan Kostic @dmkostic]

@marchiesa We cannot promise sunsets and Aurora sightings like these, but we can promise an exciting research environment and lots of academic freedom!
3/3

[Dejan Kostic @dmkostic]

More details about the project are available here:
https://t.co/4vCJB9J2bU
My colleague @marchiesa will be the co-advisor for these PhD positions
2/3