Dan Sylveste
LightNews LightNews
banner
hellblazer.com
Dan Sylveste
@hellblazer.com
180 followers 310 following 140 posts
Friend to monsters, companion to bees
Posts Replies Media Videos
hellblazer.com
An Empirical Study of Harness Design for Coding Agents

arxiv.org/abs/2609.20804

Sweet
Coding harnesses shape how autonomous coding agents translate model capabilities into long-horizon software-engineering performance, yet existing work typically evaluates harnesses as monolithic systems, leaving the effectiveness of individual components unclear. To enable component-level comparisons, we study this question with a lightweight coding harness whose execution loop is fixed while three components are varied: planning, action space, and context management. Across four models evaluated on SWE-Bench Verified and Terminal-Bench 2.1, we evaluate 176 matched settings spanning five context-management strategies, four context-window budgets, and targeted ablations of planning and action space. We find that: (1) Context management becomes increasingly valuable as the context-window budget tightens, with most of its benefit coming from preventing context-overflow failures. (2) Staging rule-based elision before LLM-based summarization provides the strongest overall efficiency among the context-management strategies, whereas making elided content recoverable adds machinery that models rarely use and yields no accuracy gain. (3) Planning shifts from an accuracy scaffold for weaker models to a cost saver for stronger models, with little change in accuracy. (4) Predefined tools improve performance for models with weaker bash proficiency, whereas bash-capable models can operate effectively with a bash-only interface and achieve substantially lower cost, especially on command-line-centric tasks. Trajectory-level analysis explains these effects: context management extends execution trajectories without substantially altering agent behavior, planning changes where trajectories stop, and the action space changes the granularity at which code is written. These findings inform model- and budget-aware harness design and provide a modular framework for evaluating future harness components.
September 18, 2026 at 9:18 PM
hellblazer.com
this, hilariously, is actually a *mobile* threat.

this is a fully embeddable, tool using agentic AI

does not require downloading 70+GB and a couple of T to run on red hot GPUs.

Not blinding fast, not a frontier model, but.... I don't think you need frontier models to be effective.
• Measured on the box with the Q2_g64 pack, then what that implies for their phone claim. What we measured (box, Vulkan, F16 KV cache): Component Size Weights on disk / in GPU memory 7.6 GB (about 2.26 bits per weight over 26.9B params) KV cache about 58 KB per token, so 3.7 GB per 64K tokens Compute buffers and overhead about 1.2 GB Total at 64K context, 4 slots 12.4 GB dedicated Total at 128K context, 1 slot 16.1 GB dedicated Host process 0.7 GB working set, 13 GB private commit (UMA mapping) The KV cache is cheap because only 16 of the 64 layers are full attention (the rest are Gated DeltaNet with a small fixed recurrent state), with 4 KV heads at 256 dims. At a phone-sized context of 4K to 8K tokens the whole thing is weights plus roughly 1.5 GB, so about 9 GB with our pack. Their phone claim uses their own packing, not ours. The ternary pack is 5.9 GB at a true 1.71 bits per weight, and the 1-bit pack is 3.9 GB at 1.125, both with custom Metal or MLX kernels. Our upstream-compatible Q2_964 carries about 30 percent more bytes than the ternary pack because of the group-64 scales, and PQ2_0 sits between at 7.2 GB. So on their kernels, ternary plus a short context is roughly 7 GB and 1-bit is roughly 5 GB, which is the "about 6 GB available RAM" they quote for iPhone. That is plausible on a 12 GB iPhone 17 Pro, where an app can hold about 6 to 8 GB. It would not fit an 8 GB phone with the ternary pack. Speed on a phone is another matter. Decode is bandwidth-bound and a dense 27B has to stream every weight per token, so at roughly 60 to 70 GB/s of phone memory bandwidth the ternary pack tops out near 10 tokens per second in theory, well under the 87 tok/s they measured an M5 Max at 546 GB/s. The box sits between those at 256 GB/s and gave us 24 tok/s with upstream kernels, so the numbers line up. The phone claim is credible for footprint, given their pack and kernels. What it costs is generation speed, and what we could not verify is quality on that pack, since the …
September 18, 2026 at 4:56 PM
hellblazer.com
Holy Shitsnacks

I can run this mf on an *iPhone*
• The real answer: the ternary Bonsai we ran is functionally sound as an agentic coprocessor. It passed the full nine-task battery. Model Battery Median wall Bonsai 27B ternary (Q2_g64) 9 / 9 2.57 min Coder-Next Q4 9 / 9 0.85 to 0.90 min Qwen3.8-27B 9 / 9 2.15 min Gemma 4 26B-A4B 7 / 9 1.60 min - Battery detail: control arm, three repeats each of the interval debug, TDD implementation, and version-compare debug tasks. Every repetition passed its verify step, with three to seven tool calls per run. So it navigates a tree, edits, runs tests, and lands green, repeatedly. - Shakeout: chat, JSON-schema synthesis, tool calling, and tokenize all pass. The three fails are vision (text-only server, by design) and the Mac's rerank service being down, neither about Bonsai. - Earlier bad sign retracted: the single wrong coding answer from yesterday's no-think probe was not representative. Under the real agentic harness it does the work. It is the slowest of the four, about three times Coder-Next's wall time, which you said is not the concern. On functionality alone it sits with Qwen3.8 in the "works" tier, from a 7.6 GB file.
September 18, 2026 at 4:39 PM
hellblazer.com
When Tool Calls Succeed but Workflows Fail: Anomalies at the Agent-Tool Boundary

arxiv.org/abs/2609.15397
Abstract Al agents increasingly execute long-running workflows that externalize effects through independently supplied tools. Under retries, speculative execution, con-currency, and partial failures, the resulting external state may be inconsistent with the workflow's intended resolution: required effects may be missing or duplicated, aborted effects may survive, and committed effects may depend on provisional state that is later withdrawn. Advanced transaction models address related fail-ures, but assume that lower-level operations expose the semantics they depend on: whether an effect occurred, whether it can be compensated, staged, or safely reordered. Shared agent-tool interfaces usually do not. We contribute an effect-history model that separates events in the external world from the runtime's observations of them, and a catalog of eight recurring external-effect anomalies. From the catalog we derive the boundary capabilities required to exclude each anomaly in general, and four points where black-box tool invocation alone cannot provide a general guarantee. We then ask how much of this is ex-pressible in a widely used shared tool interface, measuring the use of the standard annotation vocabulary across 98,291 tools exposed by registered Model Context Protocol (MCP) servers. The fields are widely emitted but provide only coarse call-level hints, and none of the required capabilities is fully expressible. These results motivate reusable transactional contracts at the tool boundary.
September 17, 2026 at 10:36 PM
hellblazer.com
EvoOntology: ASelf-Evolving OntologyLayerforDataAgents

arxiv.org/abs/2609.15779
Abstract Data agents aim to fulfill natural-language instructions over heterogeneous data, including tables, files, and databases. However, data agents face a challenging agent-data gap: heterogeneous data resides outside the agent, while the agent can access it (e.g., column names and file paths) only through generic tools. Existing approaches either let agents directly explore raw data sources or inject manually constructed semantic layers into prompts. However, neither scales well to large heterogeneous data sources nor adapts to different agent behav-iors. In this paper, we introduce EvoOntology, a self-evolving ontology layer for data agents. EvoOntology encapsulates the ontology as an MCP server comprising a schema layer, a content layer, and a tool layer, enabling agents to actively query and interact with the ontology at runtime. To this end, we introduce a builder agent for autonomous ontology construction and a self-evolution loop that continuously refines the ontology through attribution-guided typed edits that are accepted only after a backbone-conditional paired evaluation. Experiments on three well-adopted data-agent benchmarks with four LLM backbones demonstrate that EvoOntology consistently outperforms strong baselines and existing semantic-layer approaches, effectively bridging the agent-data gap and enabling more effective interaction with heterogeneous data. Code - https://github.com/ruc-datalab/EvoOntology
September 17, 2026 at 10:30 PM
hellblazer.com
September 17, 2026 at 7:25 PM
hellblazer.com
how your email finds me
A WW1 field medic typing up a report on a type writer while ensconced in a trench. It is a museum exhibit
September 17, 2026 at 4:37 PM
hellblazer.com
Current "Agentic Theatre" Status
The Cartesian Theatre, but applied to Geese.
September 17, 2026 at 2:14 AM
hellblazer.com
I'm sure this stuff is old hat to folks who actually know about this stuff, but man. there's some really weird shitzu that pops up and you just kind of stare at it and say yea. that would cause a problem.

'A bare """ shared by 463 documents'

lol. lmao, even.
463 docs 1en=3 triple-quote docstring delimiter 207 docs len=44 SPDX-License-Identifier line 192 docs len=1 a single close brace 83 docs len=1 a single open brace 68 docs len=4 a bare try: • The chash question paid off better than the op did. conexus's answer: the 919-document component isn't real content sharing at all - it's welded by degenerate chunks. A bare "''" shared by 463 documents makes those 463 one connected component under any shared-chunk relation. That's the whole 73.2% blow-up, and it's why prose tops out at 18 - documents don't emit }. I verified the client half locally: there is no minimum chunk size guard anywhere. Nothing in chunker.py, md_chunker.py or pdf_chunker.py, no MIN_CHUNK_* constant in the package. filters.py's MIN_QUERY_LENGTH is unrelated. So a one-character chunk reaches the embedder and becomes a stored, searchable vector - Voyage spend and index space on content that can never be a useful retrieval hit.
September 16, 2026 at 2:44 AM
hellblazer.com
Anyways, quite nice to see coordination primitives that are reliable, useful and can leverage LLMs, rather than flop around like half formed machinery...

Nice progression if I do say so myself
A tidy little chapter all sewn up
September 15, 2026 at 9:22 PM
hellblazer.com
Spock strumming a guitar, in uniform (likely in some back alley time flip they always get roped into) singing Steve Miller Band’s hit: SOME PEOPLE CALL ME A SPACE COWBOY SOME PEOPLE CALL ME THE VULCAN OF LOVE
September 15, 2026 at 7:32 PM
hellblazer.com
Law As Data

Quite the prescient read. (2019, Santa Fe Institute)

www.sfipress.org/books/law-as...
Cover from My paper back copy. The Runes Of Power on the cover literally flow at night and I have to keep the book cover down ato keep it from lighting up the place.
September 15, 2026 at 5:20 PM
hellblazer.com
The Aliens dude from The History Challenge in tasteful black and white, in his classic phase with "I'm not into bestiality" under him in quotes
September 15, 2026 at 3:00 PM
hellblazer.com
The coming weird times are here
I REALLU WiSH I WEREN'T LiViNG to see man-made horrors beyond my comprehension RiGHT NOW!
September 15, 2026 at 1:25 PM
hellblazer.com
VALVE Valve @valvesoftware.com •7h Steam Frame is here! Our wireless VR headset + controllers, available now in the following models: Steam Frame 256GB Steam Frame 1TB Sign up now and learn more on @steampowered.com: A pair of women's fit shorts (80's style) with these words printed on the back: YOU WILL LIVE TO SEE MANMADE HORRORS BEYOND YOUR COMPAEHENSION
September 15, 2026 at 12:18 AM
hellblazer.com
wtaf
Bernie "I am once again" meme with: "I AM ONCE AGAIN ASKING YOU TO RECONSIDER THE CHOICES THAT LED YOU TO THIS JUNCTURE IN TIME AND SPACE" under him
September 14, 2026 at 11:49 PM
hellblazer.com
Two panel Tuxedo Pooh meme Panel 1: Solipsism Panel 2: Solip:System
September 14, 2026 at 10:09 PM
hellblazer.com
The beauty of the blockhouse that it really is an immutable audit record of every transaction. And while there’s a lot obfuscation can do there’s also a lot of effort required do ensure your obfuscation is doing what you think it’s doing.
Is the Trump Family Preparing to Cash Out on Its Own Crypto? President Donald Trump and his three sons have taken a step that would permit them to sell Worl... @www.washingtonsun.com
September 14, 2026 at 3:40 PM
hellblazer.com
Two panel Tuxedo Pooh meme. Panel 1: Slop System Panel 2: Solipsism
September 14, 2026 at 1:43 PM
hellblazer.com
Monitor details Status: running Runtime: 8h 49m 16s Script: nx tuple watch --instance nexus-bd
September 14, 2026 at 12:33 AM
hellblazer.com
Behold
Scott Horton @robertscotthorton.bsky.so... • 11h Trump launches vicious assault against journalist Aaron Rupar for revealing that Natalie Harp secretly accompanied him to Ireland. www.irishstar.com/culture/ente... Donald Trump lashes out at 'deluded freak' reporter for Natalie Harp attack President Donald Trump's social media team has issued a brutal message to liberal journalist Aaron R... &www.irishstar.com
September 13, 2026 at 11:22 PM
hellblazer.com
September 13, 2026 at 11:12 PM
hellblazer.com
Taxonomy Maintenance In The Wild Over Evolving Scholarly Data: Reliability, Efficiency, and Cost-Effectiveness

arxiv.org/abs/2607.09149
ABSTRACT The rapid growth of scientific publications makes scholarly taxonomies quickly obsolete. We study taxonomy maintenance in the wild, a new problem that moves beyond static construction by continuously adapting taxonomies to evolving scholarly repositories, such as arXiv, for a given research topic. We propose GIST, a robust framework for maintaining evolving taxonomies. Unlike purely LLM-centric approaches, GIST grounds structure induction in expert-curated evidence by extracting partial hierarchies from the "Related Work" sections of papers. It integrates these partial taxonomies into a unified global taxonomy in a geometric box-embedding space, where box containment encodes the inductive bias of is-a relations. To connect semantics with geometric structure, GIST learns a bidirectional mapping between word embeddings and box embeddings. For efficient incremental updates, GIST uses novelty-aware coreset selection to update the model with representative historical signals and new evidence, avoiding costly full retraining. To handle high-velocity paper streams under user-specific token budgets, GIST further combines a hypothesized concept generator with a cost-effective evidence retrieval module. Experiments on real-world ar Xiv datasets show that GIST outperforms state-of-the-art baselines, improving Node F1 and Edge F1 by 11.0% and 13.1% over the strongest baseline while requiring only 9.6% of its runtime and 12.7% of its monetary cost.
September 13, 2026 at 10:37 PM
hellblazer.com
Tuxedo Pooh meme with: panel 1: AGI panel 2: RSI
September 13, 2026 at 2:10 PM
hellblazer.com
Literally a supervillain
Elon Musk Spends $100 Million on Anti-Trans Ads Targeting Dems in Midterm Races Some of the candidates being targeted have already capitulated to the right on trans rights. • truthout.org
September 13, 2026 at 2:22 AM