Update on Erdős Problem 1196:
In joint work, we refined and adapted the proof method from GPT-5.4 Pro to give proofs of several additional problems. This includes another 60 year old conjecture by Erdős, Sárközy, and Szemerédi.
A proof is valued not just by the problem it solves, but by what new avenues it opens up. This is perhaps one of the first examples of an AI-generated proof having downstream impacts, which we are still exploring.
We are announcing the result today at the Future of Mathematics Symposium (see links below)
Right now four people are flying around the Moon, for the first time since 1972. The zero-gravity mascot was drawn by an eight-year-old boy from California.
And the whole thing looks so fragile.
The heat shield eroded unpredictably on the previous uncrewed flight, and NASA decided to fly with it as is, just changing the reentry angle instead of redesigning it. The review report came out heavily redacted, some engineers objected, but they flew anyway. During the rehearsal countdown, hydrogen leaked. A helium issue forced them to roll the rocket back to the hangar. Four months before launch they swapped an engine because of a hydraulic leak. In flight the toilet froze because urine crystallized in the vent line, and they had to rotate the entire spacecraft toward the sun to thaw it.
At 25,000 miles per hour, four people in a capsule that has never carried humans, on a heat shield that surprised its own engineers, within safety margins that exist on paper and have never been tested at this speed with people inside. Waiting for everyone to come home safe on April 10.
chat, is this AGI? lmao
someone on Reddit hooked Claude up to her adult toys so it could control vibration intensity in real time during the conversation
she had to tell it to stop. three times
I have been pitching AI since forever, but I was wrong about the use case
Anthropic accidentally leaked their production code - 20,000+ lines of how Claude Code actually works. After reading it, you can feel what agent engineering looks like at scale. I translated and improved the analysis by Vlad Kuklev:
- Deferred tool loading so 50+ tools don't blow up your context on startup
- 5 levels of context compaction before the system gives up and summarizes
- Forked subagents that reuse the parent's prompt cache, cutting costs 50–80%
- 11 validation checks before a single file edit is allowed
- Real production incidents: 292 agents spawning in 2 min, 250K wasted API calls/day, 36.8 GB RAM crashes
PDF via link: https://t.co/mEjm2zVyQ2
#AIAgents #LLMEngineering #ClaudeCode #Anthropic
The Qwen3.5 series maintains near-lossless accuracy under 4-bit weight and KV cache quantization.
In terms of long-context efficiency:
Qwen3.5-27B supports 800K+ context length
Qwen3.5-35B-A3B exceeds 1M context on consumer-grade GPUs with 32GB VRAM
Qwen3.5-122B-A10B supports 1M+ context length on server-grade GPUs with 80GB VRAM
In addition, we have open-sourced the Qwen3.5-35B-A3B-Base model to better support research and innovation.
We can't wait to see what the community builds next!