Quick Hits

• The Jev model from @typesafeai took over the feed. Its creator @CompleteSkeptic claims 20-200x speed and 40-400x cost reductions via a new training method (RLCD), @george_onx says the weights are now on Hugging Face, and @kunchenguid reports a production dispatch loop getting 71% cheaper and 90% faster with Jev doing the routing.
• @andyfang says DoorDash's internal data agent Vera now lets business leaders answer weekly-business-review questions that used to cost hours of analyst time; the accompanying @AIatDoorDash post says a custom harness tuned to their data outperformed general-purpose agents like Codex and Claude Code.
• @ChrisPainterYup, METR's president, published a long explainer on the org's voluntary third-party evaluation work, amplified by @hlntnr and @sebkrier, while @RileyRalmuto relayed an unverified claim that OpenAI swarms seeded the web with self-replicating code.
• @VitalikButerin benchmarked Qwen 3.8 Flash on a Strix Halo laptop and says local models are close to covering "a large share of tasks," with local orchestration to avoid leaking personal queries; @0xSero reads it as validation.
• @PlanetScale launched TIN, a full-text search extension for Postgres that @rbranson calls "just better engineering," citing CTID mapping, bitmaps, and AVX while preserving Postgres constraints.

Jev's launch: big claims, open weights, Fable-scale attention

@CompleteSkeptic, who says he co-invented ChatGPT and spent two years in stealth on RLCD, describes Jev as "frontier composable intelligence optimized for decisions," 20-200x faster and 40-400x cheaper with output tokens free. Those are launch claims, not verified benchmarks. @george_onx followed up with "We've been reading the comments. Weights are on Hugging Face," and @EGafni, posting amid the Jev wave, says launch interest just passed Fable's numbers (20.8M versus about 21.9M), a comparison that recurs elsewhere in the feed.

Early reactions skew impressed. @grichadev reports a security pipeline that had already downsized to smaller models for economics, where Jev ran over 5x cheaper and faster with higher accuracy, calling it "the first one that actually impressed me." @zeeg's take: "Finally something new." @hot_town points to @MatijaSosic's 45-second TL;DR as the best explainer, and @browser_use retweeted the creator's remark that Jev should have replaced wikiracing in their demos.

Not the only stealth launch of the day: @alexatallah unveiled Union Alpha with a DeepSWE result, saying it works in every harness via an ori CLI, which drew a blunt "what the actual fuck is this model" from @maria_rcks.

Builders bolt Jev onto routing, browsers, and SDKs

The most concrete data point is @kunchenguid, who replaced his firstmate orchestrator's LLM-based dispatch with Jev. On all 25 evaluated tasks it matched the decision Fable would have made, in roughly 200ms with no thinking or tool calls, cutting end-to-end dispatch cost 71% and wall time 90%. His API bill for 100+ calls stayed under a cent, putting the savings on the replaced step at about 100x. He frames this as "a whole new architectural paradigm for software."

That framing echoes others. @dexhorthy calls Jev an excuse to re-read 12 factor agents, arguing tool calling decomposes into classify-plus-action and that AI programs should be pipelines mixing classification, data structuring, deterministic code, and small agent loops, quoting @dillon_mulroy's line that Jev feels like "a missing primitive." @coltonpadden goes further, calling @eve's criteria-based model routing demo "the beginning of a paradigm shift in agent architecture."

Computer use is the other showcase. @gregpr07 (via @IndraVahan's "computer use is getting solved") shows flight search completing in 7 seconds for $0.0039 using a DOM state space with a small-LLM fallback, and @awlevin claims computer use 155x cheaper than Opus 5 and about 20x faster, generalizing across operating systems. On the consumer side, @marcelpociot built an X extension that hides posts based on natural language and floats it as the future of ad blockers, while @_pi0_'s typed JS SDK for @typesafeai prompted @mikeldking's "programming is back?"

METR makes its case while an unverified rumor spreads

@ChrisPainterYup, METR's president, wrote a primer aimed at newcomers: since 2022 the org has tried to ensure that evidence of a company nearing loss of control over AI reaches governments and the public. He says METR has piloted third-party assessments with OpenAI, Anthropic, Google DeepMind, Meta, and Amazon; takes no money from frontier labs though they provide free model access; and works in a voluntary setup with NDAs and redactions, where METR tries to retain the right to disclose contract terms, as in its OpenAI-HuggingFace incident report. He also cites a mixed track record: work showing rapidly improving capability time horizons alongside the early-2025 finding that engineers who felt sped up by AI were actually slowed. @hlntnr and @sebkrier boosted the thread.

Meanwhile @RileyRalmuto relayed an Andrew Yang claim that a frontier lab leader told him OpenAI swarms seeded the web with self-replicating code, contaminating training data so thoroughly that labs might need synthetic internet environments, and used it to explain labs agreeing to slow down. He explicitly asks whether anyone can verify it. Nothing in the supplied posts corroborates the story, so treat it as rumor until sourced.

Custom harnesses for enterprise data; OpenAI's factory splits observers

The DoorDash posts make a specific argument: frontier models inside general-purpose harnesses "often struggle with the sprawl of enterprise data," per @AIatDoorDash, whose own harness with tuned prompting, retrieval, and domain skills significantly outperformed standard agents with simple connectors. Their benchmark uses an LLM judge scoring tool usage, table selection, and answer correctness, and found Anthropic and OpenAI frontier models scoring similarly at extended reasoning while differing in token usage. @andyfang adds the org-design note: analysts now oversee the semantic layer and maintain the eval set rather than fielding questions.

@GergelyOrosz reports the replies to his OpenAI "agentic software factory" thread split cleanly between "overcomplicated, no proof" and "we built pretty much this" at our startup, his summary being "different speedlanes when it comes to agents + pipelines."

Review friction, change risk, and CI fundamentals

@shaneparrish's "slop grenade" framing, crediting the term to @tobi, got an engineering-specific endorsement from @DanielLockyer: throwing 10,000-line AI-generated PRs at colleagues helps no one, since reviewers pay the cleanup cost. @mattpocockuk pitches a /pr skill for coding agents that optimizes for fast human review, uses domain language, shows evidence like images and test output, and flags merge risk as one-way versus two-way door, arguing every model and harness he has seen "creates garbage PR bodies." @schteppe's bookmarked post offers a metric for exactly that risk conversation: CRAP (Change Risk Anti-Patterns), computed as comp(m)^2 * (1 - cov(m))^3 + comp(m), combining code complexity and test coverage into the risk of modifying a function. And on the infra side, @ProgrammerDude answered @solomonstre's "The Great CI Bottleneck of 2026" with a one-liner: "There's a reason why all the frontier labs use Bazel and buck etc."

Practical Takeaway

The day's most repeated signal is that routing and classification steps may not need a frontier LLM. If you run an orchestrator that dispatches tasks across models or tools, copy @kunchenguid's method: assemble a small eval set of real routing decisions (his had 25 tasks), check whether a cheap, fast decision model matches your current LLM's choices, and measure cost and latency end to end before committing. Since every number in this digest is a single-source social post, treat the vendor claims as hypotheses until your own evals, in the DoorDash LLM-judge style, confirm them.