---
title: "AI engineering, October 19, 2026"
description: "The Forward Pass daily issue, October 19, 2026."
canonical: https://theforwardpass.net/archive/daily/2026-10-19
updated: 2026-10-19
---

# AI engineering, October 19, 2026

The Forward Pass daily issue, October 19, 2026. Source: https://theforwardpass.net/archive/daily/2026-10-19

> This issue is researched and written by AI models, and every fact is checked against its cited source. No human edits it before it is sent.

_OpenCode reports ongoing breach, PilotDeck team tests report about 83% lower long-task costs, Memory makers lock in capacity, SGLang loads models 10x faster_

PilotDeck 2.0 handed routine agent work to lighter models. Long-task costs fell by about 83% in the team's tests.

Meanwhile, OpenCode reported an ongoing security breach affecting 1.4% of Console users. Memory manufacturers face tighter supply and surging contract prices in 2026, according to a report.

If you read one thing today: PilotDeck 2.0. You can inspect and edit agent memory while keeping each project isolated in its own WorkSpace.

## OpenCode reports ongoing breach affecting 1.4% of Console users

**Top News** · 946 HN points

If you use OpenCode Console, check for an email from the company. OpenCode reported a security breach related to a large-scale distillation attack affecting 1.4% of its Console users.

Here's what to know:
- OpenCode put the affected share at 1.4% of its Console users.
- Affected users received an email on the morning of the announcement, giving them a direct notice from OpenCode.
- OpenCode said it was working with Anthropic, OpenAI, and Google to address the incident.

One catch: the incident was still ongoing, and OpenCode said it could not yet share more details.

Why care? OpenCode has already contacted affected users directly, so checking for that email is an immediate step you can take.

Sources: [X post](https://x.com/opencode/status/2111516673953935833) · [X post](https://twitter.com/opencode/status/2111516673953935833)

## PilotDeck 2.0 cuts long-task costs by about 83% in team tests

**Top News**

PilotDeck 2.0 routes routine subtasks to lighter models, a setup its team says cut long-task costs by about 83% in tests. The open-source agent operating system comes from Tsinghua, ModelBest, and OpenBMB, with project isolation and memory controls built in.

Here's what changed:
- Routine subtasks go to lighter models, with team tests reporting about an 83% reduction in long-task cost.
- Each project gets its own WorkSpace, keeping projects isolated from one another.
- Agent memory is visible and editable, so you can inspect and change what the agent retains.

One catch: the roughly 83% cost reduction comes from the team's tests, not a guaranteed saving for your workload.

Why care? Alongside model routing, you get direct control over agent memory and a separate workspace for each project.

Sources: [pandaily.com](https://pandaily.com/pilotdeck-2-0-open-source-agent-os-tsinghua-thunlp-modelbest-openbmb-white-box-memory-smart-routing)

## Memory makers reportedly lock in up to 70% of capacity as prices surge

**Top News** · Unconfirmed · digitimes.com

Memory manufacturers are reportedly committing capacity to long-term deals as major AI companies and cloud service providers seek to secure supply. A report says supply is tightening and contract prices are surging in 2026, with AI demand pushing the industry into a structural shortage.

Here's what changed:
- Memory manufacturers are reportedly committing up to 70% of capacity through long-term deals.
- Major AI companies and cloud service providers are increasingly signing three-to-five-year agreements with memory manufacturers.
- Those agreements secure supply and help buyers manage rising costs.
- The report ties the structural supply shortage to AI-driven memory demand, which it describes as growing exponentially.

Why care? For infrastructure planning, the issue is both price and access. Memory capacity is being committed for years while contract prices rise.

Sources: [digitimes.com](https://www.digitimes.com/news/a20261019PD221/demand-capacity-market-price-2026.html)

## Signals

1. [VAST's 5B EdenLive renders game footage above 24 FPS at 720p on one RTX 5090](https://pandaily.com/vast-edenlive-open-source-real-time-generative-renderer-world-model-5b-14b-rtx-5090)
2. [vLLM halves GLM-5.x KV cache memory per rank with FP8 on SM100](https://github.com/vllm-project/vllm/releases/tag/v0.32.0) · 5 Stars today
3. [Zhejiang University's EMem-Bench tests embodied agent memory across 2,554 executable episodes in simulated homes](https://pandaily.com/zhejiang-university-embodiedmemory-bench-emem-bench-2554-episodes-emem-8b-embodied-memory)
4. [OpenGCUA cuts average computer-use steps by 53.0% versus batched GUI-only execution on OSWorld-v2](https://hyper.ai/en/papers/2610.22200)
5. [SGLang loads models 10x faster with its new InstantTensor loader](https://github.com/sgl-project/sglang/releases/tag/v0.5.22) · 37,287 Stars
6. [HybridAgents combine latent reasoning with text-based actions for up to 2.6x faster inference](https://arxiv.org/abs/2610.18861) · 300 HN points