Tom

How to generate value with multi-agent AI systems? In Medicine, Pharma, Finance or Industrial environments.

[Tom @data_topology]

New Performance data on #mcp powered #Agents: multiple LLMs fail real world test with multi MCP server calls. Less than 30% success rate.

New insights and LLM - MCP Benchmark, "MCP Universe", by #Salesforce.

See if your #LLM is performing acceptable:
https://t.co/F3fA9QnAmP

[Tom @data_topology]

The new open-weights models by #OpenAI : Have they been worth the wait?

Do they point the way to GPT-5?

Watch their performance on a complex reasoning test:

https://t.co/Iobj5tBGIY

[Tom @data_topology]

Save to say, that after watching this test, Horizon BETA on OpenRouter is not a cloaked #GPT5 /GPT-5 from OPENai.

https://t.co/alJ4NhPk4b

[Tom @data_topology]

@JustinLin610 Causal reasoning capabilities of the thinking version of Qwen 3 2507 are really impressive. My YouTube video is already live. Congratulations 🎉

https://t.co/ttzDf44T4l

[Tom @data_topology]

@sophiamyang @MistralAI Causal reasoning of Magistral Medium is impressive, especially at 10x speed. I compared it to o3 and Claude Opus 4 in my YT video. It found a solution, where R1 failed. Smile.

https://t.co/iu4Iw1dsUM

[Tom @data_topology]

@arthurmensch Magistral causal reasoning performance is impressive. I compared it to o3 in my latest YouTube video:

https://t.co/iu4Iw1dsUM

[Tom @data_topology]

The new DeepSeek R1 0528 tested for it's causal reasoning performance.

Can open source #DeepSeek R1 0528 beat OpenAI's o3 Monster? Or at least be equal to o3?

Watch live the new reasoning excellence of DeepSeek R1 0528

https://t.co/UjOw6D8nvY

[Tom @data_topology]

@chrisgorgo Improve causal reasoning and logic thinking. Fails my logic puzzles and my elevator test.

[Tom @data_topology]

New independent coding benchmark data for Sonnet 4: worse than R1 in coding.
Worse than Sonnet 3.7.

What the hell is going on at Anthropic?
https://t.co/a76t0Tm8ep

[Tom @data_topology]

Vibe code a new app in 15 minutes - no cursor - no windsurf - no replit - no bolt - no lovable - just the new Google service: CREATE

https://t.co/2f11Yu95dL

[Tom @data_topology]

@deedydas Sonnet 4 fails a simple logic test. How can this model set the scene?

https://t.co/9pVWdZ1znR

[Tom @data_topology]

@decisionleader Sonnet 4 fails a simple logic test.

https://t.co/9pVWdZ1znR

[Tom @data_topology]

@rudrank Sonnet 4 fails a basic logic test.
Are you sure that you want to build on that?
https://t.co/9pVWdZ1znR

[Tom @data_topology]

@esindurmusnlp But Sonnet 4 fails a simple logic test. Why?

https://t.co/9pVWdZ1znR

[Tom @data_topology]

@minimaxir Sonnet 4 failed a simple logic test. Are base models not worth it any more?

https://t.co/9pVWdZ1znR

[Tom @data_topology]

@ErikSchluntz Sonnet 4 failed a simple logic test. So the base model is crap? https://t.co/9pVWdZ1znR

[Tom @data_topology]

@mattshumer_ Sonnet 4 failed on a simple logic test.

https://t.co/9pVWdZ1znR

[Tom @data_topology]

@sjgadler Sonnet 4 fails the simple logic test. https://t.co/9pVWdZ1znR

[Tom @data_topology]

@rowancheung Sonnet 4: Fails the Logic Test immediately. I had so high hopes ....
https://t.co/9pVWdZ1znR

[Alignment Lab AI @alignment_lab]

https://t.co/FHZwkNcoL4

One of the researchers I work with sent this in a group channel today, is this saying you can encourage grokking and skip the overfitting wait with a loss term on logits wrt entropy?
Would appreciate some opinions and thoughts here