AI engineering, October 17, 2026

Top News

Beam is Reflection's Apache 2.0, 501B MoE for teams that want to host and customize their model. You can access it through Hugging Face Inference Providers or deploy it through Inference Endpoints.

Here's what changed:

• Hosted inference runs through Hugging Face Inference Providers via Baseten, Fireworks, and Together.
• One-click Inference Endpoints deployments support FP8 or NVFP4, with CLI deployment available too.
• Transformers, SGLang, and vLLM support Beam, with supervised fine-tuning and reinforcement-learning recipes available.
• Each token activates 23B parameters, and the model supports tool calling and five reasoning-effort settings.

One catch: all 501B parameters must stay in memory. Plan on four B200s for long contexts and concurrency. Two B200s hold NVFP4 weights but leave little room for activations and KV cache. H200s cannot run NVFP4 natively.

Try it: install with pip install -U "transformers>=5.20.1", then load reflection/Beam-1.0-501B-A23B with the Transformers text-generation pipeline. For hosted inference, use reflection/Beam-1.0-501B-A23B:auto.

Sources: Hugging Face

Top News · 785 HN points

Anthropic reports that Claude solved the decades-old question of how several NIST ECDSA curve seeds were generated. Public-domain code lets you reproduce the seeds and selected base points.

Here is what makes this reproducible:

• The seeds came from the C library's lrand48 random-number generator initialized with Unix timestamps, not from a hash algorithm.
• Claude identified patterns in seed bits, recovered the generator's internal state, and used lattice reduction for truncated outputs.
• Anthropic also traced base points, the starting points for curve operations, for six curves generated in May 1997 to the FreeLIP random-number generator.
• The reproduction code is dedicated to the public domain under CC0 1.0 Universal.

One catch: the B-163 seed remains unrecovered, and the base points for several NIST curves remain unexplained.

Why care? Anthropic says the reconstructed origins make deliberate backdoors less plausible and increase confidence in the curves. This is an origin reconstruction, not a cryptographic break.

Sources: anthropic.com

Top News · Reported · ctee.com.tw · translated from Chinese

AI processors are moving toward larger packages with multiple chip dies and 3D stacking. TSMC says it can integrate chips at five times the reticle size, the area a lithography tool exposes in a single shot, with strong yield.

Here's what is on the roadmap:

• A14 is planned for trial manufacturing in 2027 and volume production in 2028.
• A14 uses second-generation nanosheet transistors, a full generation beyond N2, with more customer participation than N2 at the same development stage.
• TSMC expects full-year 2nm capacity in 2027 to double from 2026, following the start of volume production in Q4 2025.

One catch: advanced-process capacity is only part of the supply picture. TSMC says packaging capacity must be considered alongside it when planning for larger AI chips.

Sources: ctee.com.tw

1. OpenAI adds custom GPT-Live voices created from text descriptions through the API
2. Cursor expands cloud agent environments to support up to 300 repositories
3. OpenMOSS releases MOSS-Image, a 5B image generator with text layout control · 26 Stars today
4. FDE-Gym tests AI agents on customer delivery, with the best resolving 14 of 95 tasks · 281 Upvotes
5. LTX-2.5 world model ships open weights with day-zero ComfyUI support
6. NVIDIA previews 2.8 trillion tokens of pretraining data across 22 Indian languages