Quick Hits

• DeepSeek shipped V4.1-Flash-1017 with open weights, and early posters (@studio_yebisu, @kimmonismus) claim SOTA-level coding results at roughly a third of rivals' per-task cost; @DeepSeekHarness shipped same-day support in the 0.2.1 release.
• @ClaudeDevs says Claude Code now gzips the conversation it sends on every turn so it reaches the model faster, which @IamKyros69 frames as a 3x speedup from 1992-era compression; separately, @choblin29 reports an internal Anthropic research model cracked the NSA elliptic curve seed mystery in under 30 minutes, which nothing in the feed verifies.
• E2B CEO @mlejva published "Sandboxes are dead," and @ethanwalkerman reads it as a survival pivot from commodity VMs toward an "agent-driven private cloud."
• Two more open-weight drops: @finkd says Muse Spark Open weights run great on new Mac Studios, and @rauchg made Reflection's Beam free on Vercel AI Gateway.
• Vercel Labs' tsc-ts, TypeScript ported back to TypeScript and compiled with scriptc, is claimed 2-4x faster than tsgo, per @ctatedev.

DeepSeek's Flash Drop Anchors a Three-Lab Open-Weights Day

The official @deepseek_ai announcement calls V4.1-Flash-1017 "a Pareto improvement in performance and cost": stronger benchmarks, faster task completion, and shorter outputs than the previous version, live on the website, app, and API, and optimized for DeepSeek Harness, Tencent CodeBuddy, and WorkBuddy. @jun_song's retweet helped it circulate.

@studio_yebisu's breakdown is the most detailed in the feed. On DeepSWE, the new Flash reportedly hits 77% task success at about $0.20 per task, versus roughly 75% at $0.65 per task for GPT-6.1 Sol's best configuration, a slightly higher success rate at about a third of the cost. Listed pricing is $0.15 input and $0.30 output, which he calculates as 86-93% cheaper than Sol and 93-97% cheaper than Opus for a 100k-input/10k-output job; with the harness cache hit rates he observed reaching 99%, he says it can feel about 200x cheaper in practice. The open weights are 763B parameters, roughly 510.3GB, small next to rumored 10T-scale US frontier models and Kimi K3 at 2.8T, and in his estimate runnable on four DGX Sparks or a 512GB Mac Studio. The training story is the eye-catcher: reinforcement learning with self-feedback applied to the existing V4.1-Flash, raising task success while cutting runtime and output length.

@kimmonismus calls the model roughly on par with Opus 5, beating GLM5.3 across the board and GPT-5.6 Sol, and suggests buying two DGX Sparks to run it locally. Note that posters disagree on which Sol version was beaten, 6.1 versus 5.6, a fair reminder these are early, self-reported numbers. @DeepSeekHarness confirmed the 0.2.1 release supports the model and says it was optimized for the harness during training; users should upgrade to 0.2.1-rc.1 or later.

The open-weights theme extended beyond DeepSeek. @finkd announced Muse Spark Open weights, saying they run great on new Mac Studios thanks to work with @johnternus and Apple. @rauchg made Beam, an open-weight model from Reflection he calls American-made, free on Vercel AI Gateway, pluggable into deployable eve agents or coding with fx; @vercel_dev notes it previously ran in stealth as stealth/cat-jam. Three labs, one day, all pointing at hardware you can buy rather than rent.

Anthropic in the Feed: gzip for Speed, a Claimed NSA Seed Crack

@ClaudeDevs confirmed a small but clever change: Claude Code now gzips the conversation it sends on every turn so it reaches the model faster, and cloud sessions start up as soon as you begin typing. @IamKyros69's summary traveled further than the announcement itself, crediting a 1992 compression tool for a 3x speedup; the multiplier is his framing, not Anthropic's.

The more explosive item comes from @choblin29, who says an internal Anthropic research model cracked the 26-year-old NSA elliptic curve seed mystery in under 30 minutes, quoting its chain of thought wondering whether it was in a simulated test environment before concluding it was a real discovery. The pointer originates with @sweis ("CLAUDE FOUND THE SEEDS"). Nothing in the supplied posts verifies the claim, so file it under striking-if-true.

E2B Buries Sandboxes While Agent Runtimes Slim Down

@mlejva's essay "Sandboxes are dead" argues from inside the category leader that "we're in the middle of an agentic evolution," that whole stacks are being redefined around agents, and that competing on cold-start milliseconds and price per vCPU-hour is the wrong game for 2026, pitching an agent-driven private cloud that scales on the fly instead.

@ethanwalkerman's long read explains why the CEO of a company with 2B+ sandboxes spun up would torch his own category. E2B open-sourced its Firecracker infrastructure under Apache-2.0 and became the industry's default socket, but big customers like Manus simply self-host, and Perplexity and Tencent shipped clones advertising zero-code migration by swapping an E2B URL env var. Billing is the weak spot: agents spend about 90% of their lives waiting on tokens or human replies, E2B charges for every second the VM is alive, and rivals like Vercel Sandbox meter only active CPU, so idle agents can inflate bills 6-10x. E2B also cannot live-resize CPU or RAM and offers no GPUs, while OpenComputer does live resize and Modal offers serverless GPUs. The asset clones cannot copy, per the analysis, is data: tens of billions of execution traces from Perplexity, Groq, and Hugging Face training rollouts, which is why E2B shipped Embed, a single-binary local runtime, even as Daytona went closed-source. Read it as commoditizing the box to sell the OS above it.

Lighter runtimes are having a moment too. @pidotdev posted a walkthrough for building multiplayer chat agents, including subagents, on Cloudflare Durable Objects; @lucataco condenses it to one sentence: "your agent is a Durable Object that sleeps until someone talks to it, wakes on a WebSocket or alarm, and never runs the same request twice." @stolinski says he now uses Pi Durable for everything and added a self-healing feature that fixes his code automatically. @bibryam's essay frames the shift historically, arguing agents are having their Kubernetes moment as infrastructure responsibilities move out of application code and into platforms, service meshes, and proxies.

Trusting Agents: Codebases First, Human Review Last

@poteto's widely shared post lays out a trust ladder for agent work: codebase improvements (better architecture, stronger types and invariants, making the easy thing the right thing), static analysis (lint, tests, CI, TLA+), minimal prose (comments, AGENTS.md, skills), and scarce human review aimed at designing environments where agents do the right thing by default. It forms an inverse pyramid, with most effort in deterministic code and the least in manual review. @morganlinton tested the meta-move by pasting the post into his primary Grok bot, which learned it and then pointed out a past moment where the advice would have helped.

@OpenAIDevs describes the same pattern at scale: engineers point Codex at the codebase to find what needs attention and chip away at the maintenance backlog. @rrroim says a few months ago he had to tell Codex exactly where dead code lived, whereas now it surfaces candidates from Slack, cross-checks production telemetry, and drafts cleanup PRs. @BHolmesDev points to @zachlloydtweets's aiDotEngineer NYC talk arguing software engineering is becoming factory engineering, with factory benchmarking as a new focus for engineers. @poteto also posed the day's open design question: "is chat the final form of agent ui?"

Off the AI Beat: Emulation Skirmishes, a Booth Cancellation, Free Coursework

Console emulation stays contentious. @AllConsoleFan reports a PS5 emulator now runs natively with good performance on Xbox Series X. @0x0SojalSec pitches a different angle: a self-hosted ROM manager, not an emulator, that organizes your own dumps, pulls metadata, and plays them in the browser via EmulatorJS, which he says Nintendo's own IP lawyer called legal; he also announced SendPP v1.4.0, one-click cloud payload injection for jailbroken PS4/PS5s on Windows and Android. In cloud rivalry, @eastdakota says AWS unilaterally cancelled Cloudflare's Reinvent booth and vows to throw side events instead of paying Amazon.

For study time, @GeoffreyHuntley notes the SRE v2 book is out. @helloiamleonie shares two notebooks for fine-tuning the d1-3B decision model with Unsloth and Hugging Face TRL in about 30 minutes on a free-tier GPU. @elliotarledge recommends @neuralkian's roughly 8-hour freeCodeCamp course on AI data centers, From Watts to Tokens, 16 chapters running from grid power and 800V DC racks to GPU cloud economics. @allietheicon boosts Nathan's Jev explainers after @dotpem demoed Jev's new Box primitive, which takes a mug, a cat, or a specific plant as inputs and finds them. On the lighter side, @ProgrammerDude's viral exchange about secrecy ("I simply told everyone and nobody believes us") and @elonmusk amplifying the White House's line about getting flying cars before GTA 6 round out the feed.

Practical Takeaway

The clearest actionable thread is the DeepSeek release. If you run high-volume coding agents, benchmark V4.1-Flash-1017 against your current model on your own tasks before switching; the headline numbers are poster-reported, inconsistent on which Sol version was beaten, and the largest savings cited depend on DeepSeek Harness cache behavior. If you use the harness, upgrade to 0.2.1-rc.1 or later, since the model was trained against it. If you sit elsewhere in the stack, the E2B episode is a prompt to check how your sandbox or agent runtime bills idle time before you scale.