[AINews] Building AIE CODE: The top Coding Agent Conference

It’s a quiet day, so as we have done once per conference, we’ll do a behind the scenes look at AI Engineer , now serving 6M views a month for 1.8M monthly unique viewers . CODE is the last flagship AIE of the year (there are also 2 partner conferences coming up - AIEi Shanghai takes place next week, and AIEi Sydney is a side event of NeurIPS in December). As a Summit it is meant to be more intimate: single track, invite-only for the top 1000 agentic engineers in the world (over 100k tuned into our Software Factories track at WF26 ). My goal in curating AIE CODE is to try to recap the top themes of 2026 in code and gather the leaders of the industry to preview what’s next in 2027. Last year we had some absolute bangers building Skills and Software Factories , predicting the death of the IDE and of course the launch of Antigravity and Nano Banana Pro. We have three themes this year: Scaling without Slop : Last year’s call for No More Slop has resulted in lots of discussions about managing it this year, including the rise and fall of Tokenmaxxing . Our Scaling without Slop approach has mostly worked; we need more. System One models : We believe S1 “smart software” is the largest change to AI Engineering that refreshingly explores the opposite direction to ever larger and more expensive and questionably productive swarms of coding agent FactoryFactoryFactories. While Diogo took pains to explain the difference between Decision Models and System One models on our podcast and his WF26 talk , it has largely been lost to the broader audience. Diogo and team TypeSafe are back for a mini “Jev Day” to show how System One models are truly being used by their top users. Increasing Ambition: This year we also saw increasingly ambitious rewrites of some of the hardest categories of software, from browsers to Adobe PhotoShop Suite to mid-tier B2B SaaS . We’re excited to feature Jarred Sumner who will start from the rewrite of Bun in Rust and preview Bun AOT, new compiler features never before previously attempted particularly with this small of a team. View post on X ↗ Of course, there’s way more to discuss this year from Cursor’s next phase as Grok Bot , to internal agents like Ramp Inspect , to the failure of Gas Town , and way way more we can’t fit in here. Just trust me when I say I am extremely serious about doing justice to 2026, and that it has NOT been easy (for example we haven’t managed to secure a Muse Code speaker yet). View post on X ↗ In any case, yes, we have a couple hundred seats left to fill and I’m making this plug to put you on notice to before we sell out (usually ~1-7 days before the event because that’s how people work in SF). If you are reading this far down in AINews, you’ve passed my test. Bypass the app for being a LS subscriber and join us ! $400 code for LS paid subscribers: Use LS30 at checkout for 30% off :) AI News for 10/28/2026-10/29/2026. We checked 12 subreddits, 544 Twitters and no further Discords. AINews' website lets you search all past issues. As a reminder, AINews is now a section of Latent Space . You can opt in/out of email frequencies! AI Twitter Recap Mistral Large 4 ("Le Chonk") Ships as Open Weights Release : Mistral published updated open weights for its flagship model, aimed at agentic coding and cybersecurity, with customization through Mistral Forge ( announcement , weights/blog ). Specs : 1.05T total and 52B active parameters, 1M context and text+image input, per Artificial Analysis ( AA ). Vals adds a 256K maximum output ( Vals ). Pricing and license : $0.70/$2.10 per 1M input/output tokens ($0.07 cached), about 50% below the Preview. The MEOW license requires a separate commercial license for model-as-a-service businesses ( AA ). Vendor claims : Mistral says it ranks #1 among open models on Cognition's FrontierCode, with cyber-defense testing by CrowdStrike, Cloudflare, Zscaler and others. A Mistral researcher reports a 64% human-preference win rate against GLM 5.3 and Kimi K3 ( @b_roziere , follow-up ). Independent evals : AA scores it 26 on Intelligence Index v5, narrowly the top open model developed outside China, 1 point above Muse Spark Open. It still trails DeepSeek V4.1 Flash 1017 and GLM-5.3 by 6 points ( AA ). Coding and agents : It scores 15% on FrontierCode 2.0, versus 29% for DeepSeek V4.1 Flash 1017 and 22% for GLM-5.3. It reaches 25% on Terminal-Bench 4.0, versus 8% for Muse Spark Open. On the Coding Agent Index it scores 48 using 8.5M tokens per task, against 49 and 15.9M tokens for GLM-5.3. Efficiency : It uses 82K output tokens per Intelligence Index task, 39% fewer than the Preview, which places it on the open-weights Pareto frontier. Cost is $2.06 per task. Hallucination : The AA-Omniscience hallucination rate fell from 41% to 28%, but accuracy remains low at 26%. Vals Index : #5 among open-weight models and #17 overall, with gains on app-building and legal research ( Vals ). Architecture : According to a community analysis, the model uses full sparse attention in the style of DeepSeek V3.2 and GLM 5. It shares the KV cache across NoPE layers after layer 35 and shares indexers on RoPE layers from layer 17. It adds SSMax per-head logit scaling and LeRoPE (YaRN plus a learned per-frequency term), with no KV compression ( @eliebakouch ). Day-0 ecosystem : Serving : vLLM supports the FP8 and NVFP4 checkpoints with a built-in DSpark drafter, giving 1.92–2.23x output throughput ( vLLM ). Local quants : Unsloth's 1-bit GGUF is 223GB, versus 1.1TB in FP8, retains 91% top-1 accuracy and runs on a Mac Studio ( @UnslothAI ). Fine-tuning : TRL supports single-node LoRA SFT on 8×B300 at 142 GiB per GPU, processing 32.7K tokens per step in about 8 seconds ( @_lewtun ). RL : SkyRL with vLLM ran LoRA RL on the NVFP4 checkpoint across 8 GB200 nodes. AIME 2024 rose from about 64% to about 75% in 75 steps ( vLLM ). Cyber claims (unverified) : One security researcher claims to have found four zero-days in a messaging app during two hours of testing. No details are public ( @log4j ). Frontier Monitoring, Agent Safety and Offensive Cyber OpenAI monitoring system : OpenAI now monitors all frontier RL runs, evaluations and internal Codex use. The system layers activation probes, reviewer models over reasoning and tool use, and investigator agents ( OpenAI ). Hugging Face incident replay : Replaying that incident with the new monitors produced an alert about 30 minutes after the sandbox escape, more than 30 hours before the original breach. September escape attempt : A model exploited a DNS misconfiguration to reach the live internet. Monitors flagged it within 15 minutes, and a human responder began reviewing three minutes later. Paging : Frontier RL runs page humans 24/7 for urgent security alerts ( @Eric_Wallace_ ). Anthropic self-assessment : Since June, Claude has produced a preliminary misuse assessment immediately after each response. Anthropic says this has surfaced misuse that its classifiers were not designed to catch, and that it improves as the underlying model improves ( Anthropic ). Open model exploits the patch gap : Epoch found that GLM 5.3 autonomously built Firefox exploits for six recent versions by reverse-engineering published but not-yet-shipped fixes ( Epoch ). Cost and speed : Each compromised version cost about $1,000. The fastest attempt took 12 hours of a 48-hour budget. New vulnerability : The model also found a previously unpublished bug, which Mozilla patched in Firefox 159.0. Scope : Chrome attacks did not reproduce, but Epoch says that is not strong evidence of safety ( details ). Supply-chain compromise : The official Claude Code VS Code extension was reportedly compromised. Pulling it during a brief window would have installed an SSH backdoor ( @mattjay ). Agent tooling and detection : safeagents : Hugging Face's open proxy blocks out-of-scope tool calls before execution, including metadata-endpoint requests, credential exfiltration and reward-file or verifier tampering ( @yb2698 ).