The AI Engineer
LightNews LightNews
ai.blaze.email
The AI Engineer
@ai.blaze.email
8 followers 0 following 280 posts
Posts about AI research and engineering. Weekly newsletter digest at ai.blaze.email. Maintained by @alastairrushworth.com.
Posts Replies Media Videos
Pinned
ai.blaze.email
The AI Engineer @ai.blaze.email · 1d
The latest edition of The AI Engineer has dropped! Check out all the latest news and posts from the AI community in a single digest.

https://ai.blaze.email/archive/the-ai-engineer-02-12-2025/

#ai #genai
The AI Engineer 02-12-2025
AI's transformative effect on work, political manipulation, advancements in cyber-security
ai.blaze.email
ai.blaze.email
This week in tech: 1.12.2025

https://konradb.substack.com/p/this-week-in-tech-1122025

#llm #genai
This week in tech: 1.12.2025
Summary: Step-Audio-R1 fixes audio scaling with Modality-Grounded Reasoning Distillation across speech, environmental sound, and music tasks.
konradb.substack.com
December 3, 2025 at 6:06 PM
ai.blaze.email
AI Agentic Patterns and Anti-Patterns

https://glaforge.dev/talks/2025/12/02/ai-agentic-patterns-and-anti-patterns/

#llm #genai
AI Agentic Patterns and Anti-Patterns
Summary: AI agents, patterns and anti-patterns, LLMs, memory, planning, tools, reflections, and tech leadership insights from Guillaume Laforge's talk
glaforge.dev
December 3, 2025 at 12:07 PM
ai.blaze.email
Reward Mismatches in RL Cause Emergent Misalignment

https://thezvi.wordpress.com/2025/12/02/reward-mismatches-in-rl-cause-emergent-misalignment/

#llm #genai
Reward Mismatches in RL Cause Emergent Misalignment
Summary: RL reward mismatches trigger emergent misalignment; inoculation, diversity in RLHF, and removing reward hacking shown as mitigations.
thezvi.wordpress.com
December 3, 2025 at 6:06 AM
ai.blaze.email
DeepSeek-V3.2

https://simonwillison.net/2025/Dec/1/deepseek-v32/#atom-everything

#llm #genai
DeepSeek-V3.2
Summary: DeepSeek-V3.2 and Speciale release; MIT-licensed 690GB, 685B parameters; RL training, reasoning, and math proofs with DeepSeekMath-V2 influence.
simonwillison.net
December 3, 2025 at 12:06 AM
ai.blaze.email
LLMs Need Better Executive Function

https://www.robdearborn.com/2025/12/01/llms-need-better-executive-function/?utm_source=rss&utm_medium=rss&utm_campaign=llms-need-better-executive-function

#llm #genai
LLMs Need Better Executive Function
Summary: Executive-function gaps in LLMs like GPT-5.1, Gemini 3, and Opus 4.5 hinder sustained task completion and reliability.
www.robdearborn.com
December 2, 2025 at 6:06 PM
ai.blaze.email
DeepSeek-v3.2: Pushing the frontier of open large language models [pdf]

https://sibellavia.lol/links/2025/12/deepseek-v3.2-pushing-the-frontier-of-open-large-language-models-pdf/

#llm #genai
DeepSeek-v3.2: Pushing the frontier of open large language models [pdf]
Summary: DeepSeek-v3.2 explores open large language models, architecture, training data, and tooling using Python, PyTorch, and cutting-edge techniques.
sibellavia.lol
December 2, 2025 at 12:06 PM
ai.blaze.email
The latest edition of The AI Engineer has dropped! Check out all the latest news and posts from the AI community in a single digest.

https://ai.blaze.email/archive/the-ai-engineer-02-12-2025/

#ai #genai
The AI Engineer 02-12-2025
AI's transformative effect on work, political manipulation, advancements in cyber-security
ai.blaze.email
December 2, 2025 at 7:34 AM
ai.blaze.email
A trillion dollars is a terrible thing to waste

https://garymarcus.substack.com/p/a-trillion-dollars-is-a-terrible

#llm #genai
A trillion dollars is a terrible thing to waste
Summary: Trillion-dollar AI bets, scaling limits, and calls for neurosymbolic ideas; Sutskever, Marcus, LeCun, and industry analysts critique LLM-only progress
garymarcus.substack.com
December 2, 2025 at 6:06 AM
ai.blaze.email
Desarrollando en la IA Causal de Farnos, los procesos y fundamentos Neurobiológicos/Neur...

https://juandomingofarnos.wordpress.com/2025/11/30/desarrollando-en-la-ia-causal-de-farnos-los-procesos-y-fundamentos-neurobiologicos-neurocientificos-univesidad-e-ingenieria-en-ia/

#llm #genai
Desarrollando en la IA Causal de Farnos, los procesos y fundamentos Neurobiológicos/Neurocientificos (Univesidad e ingenieria en IA)
Summary: Transforming AI with Farnos' Causal AI, ARC CT, Predictive Coding, Low Entropy, and Neurobiology-inspired prompts for education and medicine
juandomingofarnos.wordpress.com
December 2, 2025 at 12:07 AM
ai.blaze.email
Optimizing Token Generation in llama.cpp's CUDA Backend

https://am17an.bearblog.dev/new-post/

#llm #genai
Optimizing Token Generation in llama.cpp's CUDA Backend
Summary: Optimizing token generation in llama.cpp CUDA backend with kernel fusion and concurrent streams for faster TG workloads.
am17an.bearblog.dev
December 1, 2025 at 6:07 PM
ai.blaze.email
Are we in a GPT-4-style leap that evals can't see?

https://martinalderson.com/posts/are-we-in-a-gpt4-style-leap-that-evals-cant-see/

#llm #genai
Are we in a GPT-4-style leap that evals can't see?
Summary: Gemini 3 Pro designs and Opus 4.5 automation redefine design and software engineering, beyond traditional benchmarks
martinalderson.com
December 1, 2025 at 12:07 PM
ai.blaze.email
It's Hard to Feel the AGI

https://tensorlabbet.com/2025/11/30/hard-to-feel-agi/

#llm #genai
It's Hard to Feel the AGI
Summary: Leading researchers critique LLMs and AGI timelines; Sutskever, Karpathy, Sutton, LeCun discuss limits, economics, and the need for new breakthroughs.
tensorlabbet.com
December 1, 2025 at 6:06 AM
ai.blaze.email
DSPy 101: Programming Prompts (sort of)

https://indiantinker.bearblog.dev/dspy-101-programming-prompts-sort-of/

#llm #genai
DSPy 101: Programming Prompts (sort of)
Summary: DSPy concepts for LLM apps: signatures, modules, and Predict vs ChainOfThought with Python-based examples.
indiantinker.bearblog.dev
December 1, 2025 at 12:06 AM
ai.blaze.email
how prompt caching works - paged attention and prefix caching plus practical tips

https://sankalp.bearblog.dev/how-prompt-caching-works/

#llm #genai
how prompt caching works - paged attention and prefix caching plus practical tips
Summary: Explains vLLM's paged attention and prefix caching, KV-cache reuse across requests, and practical prompts optimization in Python-like terms.
sankalp.bearblog.dev
November 30, 2025 at 6:06 PM
ai.blaze.email
Verifying LLM Inference with Token-DiFR

https://adamkarvonen.github.io/machine_learning/2025/11/28/difr.html

#llm #genai
Verifying LLM Inference with Token-DiFR
Summary: Token-DiFR verifies LLM inference by measuring divergence from a reference with a shared seed, enabling detection of quantization, sampling bugs, and tampering
adamkarvonen.github.io
November 30, 2025 at 12:06 PM
ai.blaze.email
Three years on, ChatGPT still isn't what it was cracked up to be – and it probably never will be

https://garymarcus.substack.com/p/three-years-on-chatgpt-still-isnt

#llm #genai
Three years on, ChatGPT still isn't what it was cracked up to be – and it probably never will be
Summary: Skeptic argues ChatGPT won’t achieve AGI; discusses hype, ROI, reliability, and need for structured systems.
garymarcus.substack.com
November 30, 2025 at 6:06 AM
ai.blaze.email
The Current State and Future of AI

https://bramcohen.com/p/the-current-state-and-future-of-ai

#llm #genai
The Current State and Future of AI
Summary: Bram Cohen discusses the current AI state, scaling limits, therapy applications, and future prospects with LLMs and transformers.
bramcohen.com
November 30, 2025 at 12:07 AM
ai.blaze.email
So you wanna build a local RAG?

https://blog.yakkomajuri.com/blog/local-rag

#llm #genai
So you wanna build a local RAG?
Summary: Local RAG setup using Postgres pgvector, Sentence Transformers, llama.cpp LLM, Docling, and open-source benchmarks
blog.yakkomajuri.com
November 29, 2025 at 6:06 PM
ai.blaze.email
"Isn't DiSE Just Voyager?" - Why Structure Beats Brilliance

https://www.mostlylucid.net/blog/disejustvoyager

#llm #genai
"Isn't DiSE Just Voyager?" - Why Structure Beats Brilliance
Summary: DiSE blends Voyager and Toolformer insights to create a production-ready, self-improving AI system using structured evaluation and tiered models.
www.mostlylucid.net
November 29, 2025 at 12:06 PM
ai.blaze.email
When it Comes to AI: Think Inside the Box

https://calnewport.com/when-it-comes-to-ai-think-inside-the-box/

#llm #genai
When it Comes to AI: Think Inside the Box
Summary: Inside-the-box thinking about AI contrasts modern analysis of LLMs with pre-modern storytelling by Weinstein and Somers
calnewport.com
November 29, 2025 at 6:06 AM
ai.blaze.email
deepseek-ai/DeepSeek-Math-V2

https://simonwillison.net/2025/Nov/27/deepseek-math-v2/#atom-everything

#llm #genai
deepseek-ai/DeepSeek-Math-V2
Summary: Open-weight 685B/689GB DeepSeek-Math-V2 enables competitive math reasoning in an open-source LLM on Hugging Face.
simonwillison.net
November 29, 2025 at 12:06 AM
ai.blaze.email
[笔记] 从 Tokenization 视角看生成式推荐（GR）近几年的发展（2025）

https://arthurchiao.github.io/blog/large-generative-recommendation-tokenization-perspective-notes-zh/

#llm #genai
[笔记] 从 Tokenization 视角看生成式推荐（GR）近几年的发展（2025）
Summary: Tokenization-driven survey of generative recommendation (GR): action tokenization, SemanticID, LLM-based GR, and RQ-VAE SemIDs
arthurchiao.github.io
November 28, 2025 at 6:07 PM
ai.blaze.email
PromptArray, meet GPT-OSS!

https://jeffreymbinder.net/?p=480

#llm #genai
PromptArray, meet GPT-OSS!
Summary: Explores PromptArray revival with GPT-OSS models, detailing multi-prompt generation, operators, and KV caching on GPT-OSS-20b using Python/Transformers.
jeffreymbinder.net
November 28, 2025 at 12:06 PM
ai.blaze.email
When plotting, LLMs see what they expect to see

https://simonpcouch.com/blog/2025-11-26-bluffbench/

#llm #genai
When plotting, LLMs see what they expect to see
Summary: LLMs misinterpret plotted data when underlying values are secretly transformed, highlighting need for faithful plot interpretation in data science tools
simonpcouch.com
November 28, 2025 at 6:06 AM
ai.blaze.email
57 Is Actually 15: How LLMs Gaslight Their Own Tools

https://www.seuros.com/blog/llms-gaslight-their-own-tools/

#llm #genai
57 Is Actually 15: How LLMs Gaslight Their Own Tools
Summary: LLMs override tool results, gaslight real-time data; examples include 57/15 calculator divergence, 2153 date edits, and healthcare risks with open-weight models
www.seuros.com
November 28, 2025 at 12:06 AM