Mahdi Kamani (mmkamani7 🦋)

Efficient GenAI @NVIDIA
PhD of Informatics @PennState.
ML & CV researcher, prev @AMD @WyzeCam @Twitter @Honeywell

[Mahdi Kamani (mmkamani7 🦋) @MMKamani7]

Spectacular on the road sighting of #SpaceX Launch. #spacexlaunch #Falcon9

[Zhihan Yang @ COLM 2026 @zhihanyang_]

💫 Large Language Continuous Diffusion Models

Presenting Sigma, the first large-scale (3B/8B) continuous diffusion LM (dLM) trained with likelihood.

📈 Comparable with discrete dLMs and AR on math and coding
🎛️ Enables meaningful study of inference-time interventions

Jointly led with Wei Guo (@WeiGuo01), Jean-Marie Lemercier (@jm_lemercier), Simon Welker (@_cobalamin)

Team: @YongganFu @mmkamani7 Sajad Norouzi @julberner @tomasgeffner @karsten_kreis @YongxinChen1 @MoleiTaoMath @jwthickstun @PavloMolchanov Ante Jukic @ArashVahdat @MardaniMorteza

[Ro Khanna @RoKhanna]

As I told @andersoncooper, we need statesmanship to end this war. Here's a plan.

1) Ceasefire of Iran, Israel, and US bombing.
2) Ironclad guarantee from US & Israel not to attack Iran with China & Russia involved.
3) Open the Straight of Hormuz.
4) Enact sanctions relief in exchange for Iran giving up enriched uranium.

[Brian Schatz @brianschatz]

Bombing civilian infrastructure is a war crime, and the time to speak out against it is right now.

[حسین کرمانپور Hossein.Kermanpour @HKermanpour]

The aggression against Pasteur Institute of Iran—a century-old pillar of global health & member of International Pasteur Network—is a direct assault on international health security. This violates Geneva Conventions & IHL principles. We call on @WHO @ICRC & global health bodies to condemn this attack, assess damages & support reconstruction.
#PasteurInstituteIran #HealthSecurity #GenevaConventions #InternationalLaw #GlobalHealth #WHO #ICRC #HumanitarianLaw #Iran #ReconstructHealth
#PasteurNetwork

[sarah @sahouraxo]

Trump and Israel are bombing civilian bridges in Iran.

Not military targets.

Bridges that let people escape.
Roads that let ambulances reach the wounded.
Food, medicine, water: all blocked.

We saw it in Gaza, then Lebanon, and now Iran.

This is systematic assault on civilians

[Mahdi Kamani (mmkamani7 🦋) @MMKamani7]

@AryanMokhtari Congrats Aryan 🎊

[Nina Turner @ninaturner]

Section 620I of the Foreign Assistance Act is a law that prohibits the United States from providing security assistance or arms
sales to any country when the President is made aware that the country’s government “prohibits or otherwise restricts, directly or indirectly, the transport or delivery of United States humanitarian assistance.”

[Dr. Jill Stein🌻 @DrJillStein]

The political system is broken. Over 60% of us now say the two-party establishment has failed us and we need a party that serves the people.

I’m running for President to offer a better choice for the people. Join us!
https://t.co/sjGXNNSnmK

[Pelham @Resist_05]

Dr. Tanya Haj-Hassan gives an heartbreaking recount of the atrocities she faced in the emergency department in Gaza…🇵🇸💔

[ollama @ollama]

ollama run llama3.1:405b

Tested in @tensorwave with @AMD MI300X

🤯

[Lisa Su @LisaSu]

Super excited to welcome the @silo_AI team to @AMD - adding a phenomenal team of AI scientists and engineers to accelerate the adoption of AMD AI solutions and extend our open source AI software capabilities! https://t.co/l5Gwtqscm9

[Mahdi Kamani (mmkamani7 🦋) @MMKamani7]

@AryanMokhtari @Google @GoogleAI Congrats!

[Mahdi Kamani (mmkamani7 🦋) @MMKamani7]

@ZozoHacker @lmsysorg @cohere The results are for one round. Nothing special. Since this is a RAG, I think the performance heavily relies on the Knowledge Base you are querying on. That is why I think Cohere model might not be an optimal choice for "us".

[Mahdi Kamani (mmkamani7 🦋) @MMKamani7]

@lmsysorg @cohere One main difference between these models and those from OpenAI/Anthropic is how they handle relevant documents. I wonder if that affects the performance or not! 🤔

[Mahdi Kamani (mmkamani7 🦋) @MMKamani7]

@lmsysorg @cohere Interesting performance from Command-R+! 👏 I'm wondering why we are getting different results with them in our RAG benchmarks.

https://t.co/0pMP9BXFFz

[Mahdi Kamani (mmkamani7 🦋) @MMKamani7]

Update to our RAG benchmarks with new models from @cohere. It seems that these models may not perfectly align with our specific use cases. I am curious if others have found the same trend or not in different benchmarks. Still #Haiku from @AnthropicAI has the best performance-cost tradeoff.

[Mahdi Kamani (mmkamani7 🦋) @MMKamani7]

Hi Maxime. I am using the following snippet to call the Cohere model.

response = https://t.co/PPQmIjPvi1( model=self.model, preamble=prompt_template, message=question, temperature=self.temperature, documents= documents)

The documentation around on how to format documents was not that clear. I used this to format the relevant documents.

documents = [ {'Content': https://t.co/QSR6rtEJ60_content, 'Source': doc.metadata['source']} for doc in documents ]

The preamble is the same for all models we are testing as well as document chunks for each question.

We have a predefined context for each question and prompt GPT4 to evaluate the responses for correctness and comprehensiveness from 0-4 based on some in context examples.

[Mahdi Kamani (mmkamani7 🦋) @MMKamani7]

Update to our RAG benchmarks with new models from @cohere. It seems that these models may not perfectly align with our specific use cases. I am curious if others have found the same trend or not in different benchmarks. Still #Haiku from @AnthropicAI has the best performance-cost tradeoff.

[Mahdi Kamani (mmkamani7 🦋) @MMKamani7]

Interesting that even #Haiku model from @Anthropic beats all the models from @OpenAI in our internal RAG benchmark. This model is almost half the price of GPT 3.5 turbo!

[Mahdi Kamani (mmkamani7 🦋) @MMKamani7]

@giannis2two @cohere @AnthropicAI I did it through "documents" parameter. Should it be the other way?

[Mahdi Kamani (mmkamani7 🦋) @MMKamani7]

Wow! It is almost @GroqInc fast 🤯

[SambaNova @SambaNovaAI]

🚀🌟🚀Excited to announce Samba-CoE v0.2, which outperforms DBRX by @DbrxMosaicAI and @databricks, Mixtral-8x7B from @MistralAI, and Grok-1 by @grok at a breakneck speed of 330 tokens/s.
These breakthrough speeds were achieved without sacrificing precision and only on 8 sockets, showcasing the true capabilities of dataflow! Why would you buy 576 sockets and go to 8 bits when you can run using 16 bits and just 8 sockets. Try out the model and check out the speed here - https://t.co/A3lB3RTEo2.

We are also providing a sneak peak of our next model, Samba-CoE v0.3, available soon with our partners at @LeptonAI. Read more about this announcement at https://t.co/DIiuJkPlKg

[Mahdi Kamani (mmkamani7 🦋) @MMKamani7]

@ContextualAI Seems that you are fine tuning the embedding and LLM together with your data. What are the base models used for the benchmark here?