Backscroll ranked feed — 2026-07-09 01:45Z

Tenobrus @tenobrus 2026-07-08 19:54Z

swe 1.7 just launched and it's looking pretty great vs the paretos for frontier models here... and composer 2.5 is looking pretty terrible. on *cognition's* benchmark suite. meanwhile in composer's launch post, it looks pretty pareto-great on *cursor's* benchmark suite. imo this

Soren Larson @hypersoren 2026-07-08 15:02Z

plays to capture commerce from data at the edge yourself are often slower to market but they’re also more defensible unlike in vogue data sales—with only fanciful stories of value accrual—trusted sensors often command monopolies on downstream commerce, a rare trait today

Jukan @ ICML @jukan05 2026-07-08 15:02Z

CHINA’S GOVERNMENT HAS APPROVED PURCHASES OF 200,000 NVIDIA H200 CHIPS BY DEEPSEEK, BYTEDANCE, ALIBABA AND OTHERS FOR TRAINING PURPOSES.

akira @realmcore_ 2026-07-08 22:59Z

A thing I have observed in agentic programming is that people are trending towards more and more determinism In harness behavior, cost etc. By this point it seems clear that english and markdown are not the best way to define these fixtures

Esha @esha_hq 2026-07-08 15:45Z

good writing is made "al dente." we tend to value ideas in proportion to the mental work we invested in reaching them. corp slop optimizes for "clarity" yk remove every possible ambiguity, overexplain, caveat, footnote. cushion every letter so u can sleep better when who u care

Aseel Swaid @aseelswaid9 2026-07-08 19:59Z

Tripoli, Lybia, 2002

Dwarkesh Patel @dwarkesh_sp 2026-07-08 23:58Z

Seems to suggest that if it stops being the case that there's 3 labs which are all roughly equally good, competing each others margins away, the provider of the best model could probably get away with charging *a lot* more than they currently are.

Zuhayeer Musa @zuhayeer 2026-07-08 22:59Z

Mapped SWE pay across biotech companies. Benchling, Verily, and Genentech land at the top. The hub in the Bay Area is South San Francisco, and in SoCal its San Diego, and on the east coast its Boston.

Hanchen Li @lihanc02 2026-07-08 22:00Z

Combee has been accepted to COLM 2026! Since we released the paper, parallel agent workflows have become standard in coding agents. Although we do not develop production agents, we hope the paper gives insights for people exploring aggregation in parallel agent workflows!

Lazarz @Laz4rz 2026-07-08 21:31Z

I don’t know how many times I have to repeat that now you can’t disentangle the model+harness product combo. It’s one. It’s a whole. You can straight up maximize or minimize performance on a task by changing how the model interacts with the harness.

Restructuring__ @Restructuring__ 2026-07-08 23:17Z

Excuse me, what? “The original PPM disclosed a significant fee structure, including a 1.5% management fee, 10% construction management fee, 3.5% property management fee, 2% asset management fee, and a 1% acquisition fee, before any carried interest/promote.What multiple LPs and

Will Manidis @WillManidis 2026-07-08 12:32Z

I spent years getting blacklisted by “healthcare VCs” for suggesting that cerebral was running a pill mill and not exactly “saving lives” while they sold hundreds in secondary and now the founder who they exited silently is accused of yet again running a prescribing farm

Architectural Art (A-A) @4AAAAart 2026-07-08 06:35Z

Tables by Brunel*

Anjney Midha @AnjneyMidha 2026-07-08 15:55Z

psa: if your infra engineer proactively flags that your tensor core utilization sucks, promote them many exaflops are being wasted at AI labs right now as a result of treating utilization as an afterthought investors are starting to do the math and will ask you hard Qs soon

Susan Zhang @suchenzang 2026-07-03 13:52Z

books3 was absolutely necessary for llama to happen back in the day. you can even find my pathetic cleanup code for it here: https:// github.com/facebookresear ch/metaseq/tree/main/preprocessing/books3 … it was also noted to be a critical datasource for mmlu in the llama1 report, though by now we know mmlu is useless and high

Kris Luckenbach @KrisLuckPhoto 2026-07-08 22:05Z

Shorebird & Songbird Photography on a beautiful day on Kodiak Island

Rishabh Agarwal @agarwl_ 2026-07-08 18:22Z

Meta's offer was an order of magnitude higher than $1 million

joshpuckett @joshpuckett 2026-07-08 15:30Z

chicago riverwalk. as a woodworker/hobbyist architect I really marveled at the continuation of the soffits into the interior. That whole intersection between glass wall and ceiling must have been a fun detail to work out

Nathan Flurry @NathanFlurry 2026-07-08 05:00Z

wtf we're about to see a crazy infra shift i asked codex to ship a yolo task (no constraints) it defaulted to Durable Objects + SQLite ...not postgres, not aws, not docker

Aseel Swaid @aseelswaid9 2026-07-09 00:01Z

Sabuceda, Spain, 1977

Andrew Curran @AndrewCurran_ 2026-07-02 20:54Z

On the heels of reports that META is exploring a move into compute-as-a-service like xAI, Mark Zuckerberg told an internal town hall that AI agent development over the last four months hasn’t accelerated 'in the way we expected'. The race continues to narrow.

will depue @willdepue 2026-07-08 11:36Z

I feel confident that while long-range planning is a side effect of attention gradients in transformers, it is only so indirectly, and that an explicit mechanism to optimize thinking ahead would be a substantial win. I think multi-token prediction is a good start of where to

/ @gazorp5 2026-07-09 00:57Z

looks overly complex, expensive, and easy to break. generalist's approach (basically UMI-style tweezers with force sensing) seems better. cheap, dexterous, and easy to train.

Soren Larson @hypersoren 2026-07-08 18:20Z

information advantages from > Data at the Edge from deploying sensors literally onto the data generating process itself that prev businesses only informally instrumented / proxied very often constitutes a - GTM wedge - superior value prop - even disintermediation

blackstar @blackstarops 2026-07-08 23:24Z

“god, if i die young, send me to heaven. i’ve already been in hell” - a croatian soldier posing with graffiti during the croatian war of independence (1991)

Zephyr @zephyr_z9 2026-07-08 15:51Z

YTMC, CXMT, JHICC will bring around 7 cleanrooms (100k WPM each) online in 26 & 27

Podcast Alpha @PodcastAlphaX 2026-07-08 20:44Z

Jonathan Ross @JonathanRoss321 on Founders: Jensen Huang has never had a one-on-one meeting. That is not a quirk. That is how Nvidia stays the least political large organization Ross has ever worked with. Ross explained: when a leader speaks with two people separately, each

GEOFF @geoffwoo 2026-07-08 18:20Z

fission diligence: slide: reactor reality: NRC clock, HALEU supply, weld QA, turbine island, site politics which one kills the company first?

Vincent Weisser @vincentweisser 2026-07-08 17:30Z

We raised $130M @ $1B for our series A To build the open superintelligence stack for everyone Pre-training concentrated frontier AI in a handful of labs. RL changes who can build frontier AI and just works across almost any verifiable domain. We want to enable everyone to

alexey @sekachov 2026-07-08 18:10Z

adding some last-minute improvements: you can control the corner radius and the spacing between elements in your grid to your liking using the trackpad (hold ⌘ for corner radius or ⌘+⌥ for spacing, respectively) release tomorrow.

Amesh Adalja @AmeshAA 2026-07-08 11:36Z

The framing of "100% American-made medical gloves" is the wrong goal. The right goal is supply chain resilience — diversified suppliers, maintained stockpiles, surge agreements. Trying to solve a pandemic prep problem with economic nationalism got us neither gloves nor readiness

Xiangyi Li @xdotli 2026-07-08 22:15Z

audit on evals should be a pre-thought and not an afterthought imagine all the tokens wasted running on broken tasks. we should definitely create a structure to de-incentivize slop eval creators

Ivory Tang @ivory_tang 2026-07-08 19:39Z

the battle for fde's has only just begun, why hire 1 when you can acquire a bunch?

Pablo Antonio @PabloAntonio 2026-07-08 20:28Z

Tribeca apartment, NYC, 2035. Greco-futurism.

Daniel Kuntz @dankuntz 2026-07-08 01:47Z

Lark now supports haptic keyframing. For STARBOY we chose a really nice little LRA with a super snappy rise / fall time. All 550+ animations have keyframed haptics. We also randomly offset the magnitude of every keyframe, so each animation feels a bit different every time.

alchemy @alchemy_run 2026-07-08 22:23Z

We built a loop that automates Infrastructure as Code development for any cloud provider. One run: 230 Cloudflare resources, ~700 live tests against the real API, and an SDK with 1,087 patched operations spanning 94 of 114 Cloudflare services.

Lon Ligrin @LonLigrin 2026-07-08 17:59Z

To reduce whisper hallucinations in noisy/music-heavy audio where VAD fails: Use segment-level no_speech_prob. Try: Drop if >= 0.65, except tiny interjections <= 2 words, <= 0.45s, and < 0.85. Always drop if >= 0.85. No Amara anymore, thank me later

Nikesh Arora @nikesharora 2026-07-08 11:15Z

Guessing you mean consumption, unfortunately open source not contributing to the wallet (free). Presuming it includes the coding share which is a perfect case of a vertical app with memory and context well intertwined with the intelligence. IMHO that will be a primary outcome.

elie @eliebakouch 2026-07-08 14:05Z

FrontierCode v1 vs v1.1 on the same graph, better grading and web search rules seem to improve relatively equally all the models across reasoning effort. the number of output tokens/cost/tool calls are also similar

Maitreyee Wairagkar @Maitreyee_W 2026-07-08 23:44Z

New preprint: Brain2voice 2.0—a BCI that synthesizes highly intelligible voice directly from intracortical neural signals in real time. Human listeners transcribed voice with WER 5.24% (94% words correct)—an 8x improvement over our Nature 2025 results. https:// doi.org/10.64898/2026. 06.30.735633 …

murat @mayfer 2026-07-08 22:56Z

my local publicly owned monopoly couldn't handle a few concurrent pageloads so instead of adding an index to their database they hired a third-party company to manage a traffic queue i love it so much

Hugo @hugosaintemarie 2026-07-08 00:31Z

snapshot toast has 3 speeds: • quick: no progress bar • medium: progress bar • longer: progress bar + counter http:// paper.design/snapshot-exten sion …

Ethan Mollick @emollick 2026-07-08 01:37Z

Prompt: ANNALS — a living kingdom in a single file Working title behavior: the app titles itself per seed — "The Annals of Vaelmere", "The Annals of Osterholt" — because the central conceit is that you are watching a history write itself. The kingdom is the simulation; the

Jim Haseloff @jimhaseloff 2026-07-08 18:27Z

Marchantia plants grown on a large Super-NFT grow-pod using micro-perforated plastic film to supress algal growth, while rhizoids penetrate the support membrane and feed off nutrients flowing below. Grow healthily for months on a tabletop for spore production

James @architect_UX 2026-07-08 20:54Z

The Emile sofa by Christophe Delcourt for Molteni & C. Featured in "Geometry of Life," set in a Milanese palazzo.

Terrence Rohan @tmrohan 2026-07-08 18:58Z

The current top 5% of seed deals are the worst asset class in venture. They carry the loss ratio of a seed deal with the upside profile of a Series B. Seed risk at Series B prices.

Gizem Akdag @gizakdag 2026-07-08 19:39Z

Today, during Stage 5 of Tour de France, the riders of UAE Team Emirates wore helmets featuring my artwork. Here's the concept behind it: “Intelligence in Motion and Visibility at Speed” It begins with a structured city, representing human systems, infrastructure, and order. As

Hunter @StatisticUrban 2026-07-08 17:53Z

Actually, property taxes should be higher, possibly with a tax break for families with kids under 18. Efficiently distributing housing is just a much more important goal than ensuring Boomers can hang on to large homes that growing families need, simply for sentimental reasons.

Everyone says AI can't keep an illustration style consistent. So I built a whole coffee brand from one skateboarding pelican. Logo, badges, merch, etc. Every piece matches. Made in QuiverAI.

Rhys @RhysSullivan 2026-07-08 05:48Z

fable started an "adversarial review" of its own work and then triggered its own safeguards

Hunterbrook @hntrbrkmedia 2026-07-08 17:01Z

NEW: The CEO of an AI giant said again and again: "There is no China supply chain for us." Trade records, satellites, Chinese filings and even its own suppliers in China say otherwise. The truth raises an existential threat to Bloom Energy’s story. Our $BE investigation:

Ryan Brewer @ryanbrewer 2026-07-07 14:36Z

Doesn't count equity my guy

Apache Hudi @apachehudi 2026-07-08 15:00Z

The Hudi file group is the primitive that makes Lance fragments, RFC-80 column families, and wide AI tables all click into place. One record key → one file group. Indexed, versioned, concurrency-controlled. Two AI workloads land on it ↓

Herbie Bradley @herbiebradley 2026-07-08 11:27Z

Learning diverse "dialects" is critical to modeling the world well IMO, and lots of people thinking about technological change would improve if they broadened their repertoire. This is about Satya's essay on AI market dynamics, which @TheZvi bounced off.

Chris Tate @ctatedev 2026-07-09 00:37Z

Introducing Native SDK The toolkit for building native apps → Hot reload → Markup + Zig → Instant launch → macOS, Windows and Linux → GPU engine built from scratch → Built-in design system and themes → Custom components + design tokens

Aseel Swaid @aseelswaid9 2026-07-08 19:47Z

Sudan 1954

Susan Zhang @suchenzang 2026-07-08 00:48Z

the law of finite egos continues to hold: the more you can scale talent against a finite ego tolerance, the more luck you can manufacture in bounded time

Rob Palmer @robpalmer2 2026-07-08 16:33Z

TypeScript 7.0 is out It's the 10x faster Go-based port `tsc6` is for side-by-side API access --checkers sets checker thread count --singleThreaded keeps parsing/emit/checking together TS 6.0 deprecations are hard errors API access is planned for TS 7.1

shaunda devens @shaundadevens 2026-07-08 18:51Z

Hyperliquid has updated its documentation: priority fees, previously limited to aggressive IOC orders, now extend to ALO orders on testnet, allowing makers to bid for queue position. Expect both write and read priority fees to continue rising. With both sides of the book now

Kris Luckenbach @KrisLuckPhoto 2026-07-08 16:25Z

The Setup vs The Shot