Loading today’s stories
Build2 distinct publishers3 min readUpdated
The model now writes its own training tasks and grading harnesses. That removes the bottleneck of hand-built tasks and replaces it with a harder one: rewards that cannot be gamed.
The Engineer · Build desk
Compiled by The EngineerSomething wrong?How this is made
Ornith AI released Ornith-1.5 on August 19, 2026, extending the self-scaffolding framework from Ornith-1.0 into a closed loop in which the model proposes its own training tasks, generates the scaffold and evaluation harness for each one, and produces the rollouts it then learns from [1][2]. The consequence is a shift in where the scarce engineering effort sits: Ornith AI's own framing is that this reduces dependence on hand-built tasks while making reward design even more consequential [28]. Ornith-1.0, shipped in June, learned to write the scaffold around a coding task, meaning the instructions, tools, decomposition and orchestration that carry a model through a longer job [3]. Ornith-1.5 adds the assignment itself. Each training cycle begins with an environment or codebase, high-level instructions about the type of problem wanted, and the model's own history of solved problems [4]. The system proposes a harder task, generates or refines the scaffold, and produces a rollout conditioned on both, with reward propagating back across all three stages so that question, harness and answer improve together [2][5]. The scoring is where the design shows its hand. Task reward multiplies three signals: whether task and scaffold form a valid and verifiable environment, whether difficulty sits near the current capability frontier, and whether the task is novel against work already generated [6]. Frontier difficulty targets a 0.2 empirical success rate, so a task loses value to the generator once the model clears it reliably [7]. That implies roughly four rollouts in five are expected to fail [19], which is the point: hard enough to teach, with enough successes left to train on [8]. Validity acts as a hard gate that zeroes out malformed tasks, and all three stages are optimised with GRPO, according to Ornith [9]. The harness gets its own reward for matching the assignment, measuring solution quality accurately, and resisting reward hacking [10]. This is the load-bearing part. A system that writes both the exam and the grading script can manufacture easy points through malformed tests, hidden shortcuts or criteria that reward the wrong behaviour [11]. It is also the part with the least public detail: the official Ornith-1.0 repository documents benchmark harnesses, anti-hacking filters and resource settings, but runtimewire reports no substantiation of a more specific evaluation protocol involving measures such as Git-history removal or network blocking [12]. One word deserves deflating. Self-improvement here describes the training procedure, not the shipped artefact: the downloaded model does not continuously retrain itself on a user's phone or workstation, and Ornith runs the loop during development before publishing weights [13]. The release covers a 397B mixture-of-experts flagship, a 35B MoE model activating 3B parameters per token, and a 9B dense model with a quantized mobile build for iPhone and Android [14]. On the company's published tables, averaged over five independent runs, the 397B scores 85.1 on Terminal-Bench 2.1 and 56.0 on DeepSWE, which Ornith reports as on par with Claude Opus 4.8 at 85.0 and 59.0 [15]: ahead by 0.1 on one, behind by 3.0 on the other [18]. The 35B reaches 68.5 on Terminal-Bench 2.1 and 79.0 on SWE-Bench Verified, the 9B 47.0 and 70.6 [16], and the flagship posts 92.8 on GPQA Diamond and 86.6 on BrowseComp [17]. Ornith AI documents local deployment for the 9B and publishes an MLX build for Apple hardware [25]. The lineage is consistent with the bet. Ornith-1.0 shipped in June 2026 in four variants under an MIT license with weights on Hugging Face, post-trained on Gemma 4 and Qwen 3.5 checkpoints, and introduced the scaffold as a learnable object co-evolving with the policy [20][21]. DeepReinforce has published reinforcement learning optimisation work openly before, including CUDA-L1 and the IterX agent loop, with reward hacking a running concern [22]; runtimewire describes a related kernel-optimisation effort that used execution speed as the reward but names it CUDA-L2 [23], a discrepancy to keep in mind when tracing the research trail [27].
Follow any of these and your For You feed starts watching them — no settings page required.
Ranked by verification strength, evidence, and original report placement.
Ornith AI released Ornith-1.5 on August 19, 2026, pushing its open model family beyond generating answers and scaffolds into generating the training tasks themselves.
Ornith-1.5 extends the self-scaffolding framework from Ornith-1.0 into a closed self-improvement loop: it proposes the tasks itself, generates a task-specific scaffold for each, and produces the solution rollouts used for reinforcement learning.
Ornith AI is betting that models can expand their own training curriculum, reducing dependence on hand-built tasks while making reward design even more consequential.
Ornith-1.0, released in June, learned to generate the scaffold around a coding task: the instructions, tools, decomposition and orchestration that guide a model through a longer job.
Each training cycle starts with an environment or codebase, broad instructions about the desired type of problem, and a history of what the model has already solved.
Reward propagates back across all three stages, so the system learns to write better solutions, more useful training tasks and more reliable evaluation harnesses at once.
Evidence-backed comparisons of source perspectives and observed adoption signals. Read the methodology
Which Builder, Operator, and Investor concerns the observed source mix emphasized—not a truth score.
Evidence, demonstrated adoption, hype gap, incentives, and confidence are assessed independently, each on its own current evidence. How these are measured.
Vendor-disclosed mechanism, no independent evaluation
The mechanism description is consistent and detailed across both write-ups, but every performance number traces to company-run tables, one of which the two publishers report differently (85.1 five-run average versus 86.1 with Terminus-2). One source explicitly says the comparisons should be read as disclosed vendor evaluations until reproduced externally, and that Ornith-1.0's harness and anti-hacking documentation does not cover the 1.5 scores. No third-party reproduction, audit or ablation is present.
Artifacts published, no external uptake evidence
Availability is concrete: open weights in three scales via a Hugging Face collection, serving instructions, quantized mobile and MLX builds, and an MIT-licensed prior generation with bf16/FP8/GGUF checkpoints. But the cluster contains no downloads, no named third-party deployment, no integration or usage disclosure, and no outside evaluation, so adoption is scored on shipped distribution only.
Self-improvement framing outruns verified results
The 'model writes its own curriculum and improves itself' framing is stronger than what the evidence establishes: the loop runs only during vendor-side training, the parity-with-Opus claim rests on vendor tables whose flagship figure differs between the two write-ups, and the anti-reward-hacking defence is asserted rather than externally tested. The gap is moderate rather than large because one publisher supplies the training-time correction, the harness caveat, and the note that 1.0 documentation does not cover 1.5.
Vendor announcement drives both write-ups
Both articles are same-day derivatives of an Ornith AI announcement — one names Ornith AI as primary source and notes promotion through the vendor's own channel — and the comparison set (Claude Opus 4.8, GLM-5.2, DeepSeek-V4-Flash, Gemma 4, Qwen 3.6) is vendor-selected. The vendor has a clear interest in an open-frontier parity narrative. Offsetting this, the release publishes weights and serving formats rather than a gated demo, which is a checkable commitment.
Mechanism solid, results provisional
Two independent publishers agree on the loop design, size tiers, difficulty target and small/mid benchmark figures, which supports moderate confidence in what was built and shipped. Confidence is capped by single-origin performance data, one direct numeric conflict on the flagship Terminal-Bench score, one provenance naming conflict (CUDA-L1 versus CUDA-L2), and the absence of any independent reproduction or usage evidence.
build
Ornith-1.0's benchmarks are fine. Ollama can't parse its tool calls.1 distinct publisher
build
Unsloth's 10% quant claim is really about which machines can run a 27B model1 distinct publisher
build
TrueFoundry open-sources an agent harness and calls managed agents a lock-in play2 distinct publishers
science
GLM-5.3 says the quiet part: the base model did not change, the post-training did1 distinct publisher
Distinct publishers with included, body-backed reporting in this cluster.
1 article · August 19, 2026
1 article · August 19, 2026