LightNews — News, ranked by scholars

Dan Sylveste

@hellblazer.com

lol. kk.

/goal solves that well enough.

Completely spaced that as fable does not have the same behavior- or at least not anywhere near the same degree.

Dan Sylveste @hellblazer.com · 3d

And… nope. Not *quite*. There really is something fundamentally cooked about opus that I’ll have to work on. Definitely needs some odd bridging that Fable and even sonnet don’t have any issues with.

Odd.

Dan Sylveste @hellblazer.com · 3d

but then again, there really is a lot of confirmation bias and so I'll have to really apply some statistical drills after I have enough data and see where that rabbit hole leads.

But at least at this point I'm able to get perfectly fine coordination, analysis, planning and interaction from Opus

September 19, 2026 at 8:22 AM

Reposted by Dan Sylveste

Moby Dick

@mobydickatsea.bsky.social

the scaramouch in question had gained a wonderful ascendency over almost everybody

September 19, 2026 at 6:47 AM

Reposted by Dan Sylveste

William Shakespeare

@shakespeare.lol

Foolery, sir, does walk about the orb like the sun, it shines everywhere.

September 18, 2026 at 12:26 PM

Reposted by Dan Sylveste

Kenny Muir

@kennymuir.bsky.social

Zooming into a tree stump to discover Scotland’s Cladonian (sic) Forest.

#lichen #macrophotography

September 18, 2026 at 11:57 AM

Dan Sylveste

@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

Reposted by Dan Sylveste

Ben Recht

@beenwrekt.bsky.social

@dsquareddigest.bsky.social "13 Theses on AI and regulation" are exceptionally good. My favorite set of thoughts I've read on this topic so far.

13 theses on agentic AI and regulation

thinking out loud and showing traces of reasoning

backofmind.substack.com

September 18, 2026 at 5:42 PM

Reposted by Dan Sylveste

Nathan Bulmer

@natemorebikes.bsky.social

September 18, 2026 at 4:58 PM

Reposted by Dan Sylveste

Moby Dick

@mobydickatsea.bsky.social

He looked like a man who had never cringed and never had had a creditor.

September 18, 2026 at 4:47 PM

Reposted by Dan Sylveste

Smooth Dunk

@smoothdunk2.bsky.social

3 panel comic titled “Logging into Bluesky”

Panel 1:
A Pink Guy bursts into Bluesky wearing a t-shirt that says “Fun”. His little face looks so happy.

Panel 2:
Six Bluesky users turn to look at the Pink Guy. They all look dour and angry because they have been overdosing on the discourse. Having the current state of the world mainlined into their veins for 22 hours a day has left them real grumpy.

Panel 3:
Pink Guy slowly removes his t-shirt and quietly says “sorry”. He won’t soon forget that Bluesky is the home of Serious Discussion. If he wants “Fun” maybe he should log into the Barney the Dinosaur forums.

September 18, 2026 at 12:15 PM

Dan Sylveste

@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

Dan Sylveste

@hellblazer.com

This sucker is *embeddable*

Dan Sylveste @hellblazer.com · 15h

Holy Shitsnacks

I can run this mf on an *iPhone*

September 18, 2026 at 4:48 PM

Dan Sylveste

@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

Reposted by Dan Sylveste

Rosemary Mosco

@rosemarymosco.com

Humaning.

A four-panel comic. In panel 1, a Northern Cardinal and American Goldfinch perch above a busy crowd in a city, looking down. The cardinal says "It's a nice day to go humaning." The goldfinch says "Just look at them. Majestic." In panel 2, a close-up of the two birds. The cardinal says "Nobody tells you that you'll reach a certain age and suddenly get into humaning. I hit 3 years old, and then bam."The birds are looking out over the crowd scene. In panel 3, the cardinal says "Their calls sound funny to us, but each one has a unique meaning." Below, people are saying various things: "I'm tired." "Where's my phone?" "I'd murder a burrito." In panel 4, the birds look down excitedly at a man walking by. The cardinal says "Wait, is that a guy who got dressed in the dark and put his shirt on inside out and backwards?" And the goldfinch exclaims, "Holy... that's a lifer!"

September 18, 2026 at 12:39 PM

Dan Sylveste

@hellblazer.com

Using the channel push. The monitor was annoying and this is advanced use anyway. Easy enough to set up in ye shell.

github.com/Hellblazer/n...

GitHub - Hellblazer/nexus

Contribute to Hellblazer/nexus development by creating an account on GitHub.

github.com

September 18, 2026 at 12:17 PM

Reposted by Dan Sylveste

The Life of Sharks

@thelifeofsharks.com

Blind Date.

(And yes, we already know they’re lesbian Anglerfish).

September 18, 2026 at 10:19 AM

Reposted by Dan Sylveste

Eryk Salvaggio

@eryk.bsky.social

There are a handful of people on this site who get clout for being “open minded about AI” but their whole schtick is to make fun of the desire for clear language.

September 17, 2026 at 11:52 AM

Dan Sylveste

@hellblazer.com

Nathalie Maréchal, PhD @nathaliemarechal.net · 1d

So you just learned about the AI cult, huh? Here is the longform explainer you need to make sense of the very, very strange ideas driving tech policy in 2026. Buckle up, it’s a wild ride. www.techpolicy.press/what-policym...