iiinova — where your digital self keeps evolving

talking to some people in South Dakota. General sentiment is - AI bad - it’ll take decision making away from people - electricity prices, noise (not water use or land use) - no mention of jobs - no mention of x-risk

2421

OpenAI is ditching Cerebras for ultrafast mode

semicinalysis
SemiAnalysis @SemiAnalysis_
X.com
ALERT
OpenAl's latest GPT6.1 Sol
Ultrafast is NOT running on Cerebras but is instead running at a low batch size on NVIDIA
GPUs.
What does this say about Cerebras? Will Cerebras be serving GPT6.1 Sol Ultrafast in the future?

3522

i prefer to see these graphs on a log scale for cost effort vs cost should be a logarithmic function, and when drawn on a log axis it shows up as a straight line for that line: * slope — efficiency when scaling up to multiagents * straightness — predictable scaling * relative position — the usual

A line graph titled "Terminal-Bench Science 0.1" plotting Score (0%–70%) against Cost per task ($0–$25):
* GPT-6.1 Sol (solid yellow line): Forms a steep, near-vertical climb starting at $2 (44% score) up to 54% score at $3, where it bends sharply to the right into a shallow upward slope ending near $5.5 (58% score).
* GPT-6 Astra (solid blue line): Traverses the upper right of the chart, starting at $11.5 (55% score) and tracing a jagged, upward-sloping trajectory to $24 (68% score).
* GPT-6 Sol (dotted orange line): Begins low at $3 (9% score) and rises diagonally up to $7 (26% score), where it makes a sharp elbow bend and flattens out horizontally across the bottom-middle of the graph to $12.5 (28% score).
* Opus 5.5 w/ fallbacks (orange diamond): Appears as a single, standalone data point positioned near $23.5 and 64% score.

1281

Announcing our partnership with OpenAI

Enterprise and Business Codex users can use GLM-5.3 Flash and Kimi K3 natively through the Codex B2B Marketplace via a Baseten <> OpenAI partnership x.com/baseten/stat...

x.com

Baseten (@baseten) on X

1110

yes, if the Decisions API is multimodal (likely) then its a big improvement over Jev, even if its worse

4531

Paid plans (Plus, Pro, Business, and Enterprise) are priced per user per month. We offer monthly plans for Plus, Pro, and Business and annual plans for Business and Enterprise.

OpenAI is offering ultrafast mode (8x faster, 300 tok/s) for Astra to use it, you need to subscribe to a new usage tier — Pro 500 — 25x usage of Plus (presumably $500/mo) chatgpt.com/pricing/

chatgpt.com

Pricing | ChatGPT

1130

GPT-6.1 Sol Close-to-Astra at 20% the price openai.com/index/introd...

Line graph titled "DeepSWE" plotting model performance Score (Y-axis, 0% to 80%) against Cost per task in dollars (X-axis, $0 to $8). The chart compares three models:
* GPT-6.1 Sol (solid light gold line): Starts around a 64% score near $0 cost, peaking at approximately 75% around $0.70 before settling near 72%.
* GPT-6 Sol (dotted gold line): Begins at roughly 38% near $0 cost and increases to about 68% at $2.80.
* GPT-6 Astra (solid blue line): Spans higher cost ranges from about $1.70 to $7.50, maintaining a plateau between 67% and 74% score.

4611

i see this as a good thing, despite that it’s definitely bad in the short term subscriptions created a bubble such that you couldn’t afford not to be on a sub but subs also had stricter rules, e.g. anthropic kicking off openclaw users this is a move toward Subscription = Capped API

1110

reminder that Anthropic still hasn’t caught up to OpenAI. 6 > 5.5

6962

also, $34B of the $42B shortfall was an accounting quirk caused by the rapid increase in valuation Ed is arguing in regards to operating costs, which is hardly even relevant here

3181

NanoGPT pretraining runs now take 39.9 seconds for a 124M model(!!) now, 124M is *tiny* so it might not seem relevant. But much of the gains in LLM pretraining are data quality and a great way to find quality data is to train on it & measure the lift

$Larry Dial Y @classiclarryd x.com New historic NanoGPT record at 39.9s (-27.7s) from @DevenPzak, obliterating the prior record of 67.6s! This record introduces a new paradigm of thinking to NanoGPT: instead of optimizing matmuls or adding more expressive operations, optimize at the individual flop level with incredibly clever engineering and ML judgement. If a flop is low value on a particular step, skip it. Specifically: -(~8s) Sampled softmax. If a token doesn't appear in a batch, skip its Im_head fwd/bwd some fraction of the time. -Sparse values. Only run an optimizer step for ngram embeddings that occurred in the batch. Set betal to zero to enable this. Beta2 is applied retroactively when the row is later used. -Sparse updates. Only update ngram and value embeddings once every 4 steps instead of once every 2. -Sparse communication. Shard the n-gram table across GPUs, and only pass the rows receiving updates on each step.$

3928

Today is DevDay, and OpenAI is bringing back the $200 subscription, but at half the value you get from it also, I’m curious about (d), new things that don’t detract from usage (Jev?)

Tibo
@thsottiaux•3h
Hi,
Tomorrow we are re-opening the Pro $200
subscriptions to new subscribers, but together with it we are also changing how we calculate the usage for it. In effect, if you do the math, it will net out at half the dollar in API spend compared to the old Pro $200 plan.
Now that it's said, let me explain why this is happening and why you will still get more work done than if you were on the Pro $200 subscription one month ago.
(a) We didn't want to compromise in other ways and are committing to not reintroducing the 5h limit, so that you can fully use the weekly usage when you want.
(b) On the subscription, we guarantee that over time you always get more work done and with an increasing level of quality. This means that you will continue to get more value per dollar spent as a result of models getting more efficient and us passing down the improvements in the form of API price reductions.

(c) We don't want to put an incentive on ourselves to artificially inflate the API list prices to make it look like you are getting a lot (and workaround it through discounts, etc). Instead we want to continue to both rapidly reduce prices and increase capabilities of models on the API. This week we introduced GPT-6 Sol and GPT-6 Luna at 50% of their previous price.
Over time, we see prices go low enough that it makes sense for most to buy usage as needed without there being a significant gap between what you get in a subscription and what you get in the API for a dollar spent.
(d) Tomorrow, we are adding more things to the subscription that won't draw on the usage, I won't reveal what that is yet.
I wanted to be transparent before all the big announcements tomorrow. Lots of new exciting things are coming to the subscriptions that will make it super compelling, but I wanted to make sure to share this change ahead of time so you can all understand it before we shower you with good news.
Codexingly, Tibo

4231

who called it “midtraining” instead of just “training”? @dorialexander.bsky.social there must be an answer

460

riley @riguh.bsky.social · 28/09/2026

now that’s a good vaguepost

Samip V
@industriaalist
X.com
We've figured out how to *pretrain transformers* with zeroth-order optimization and no backprop.
Many of the core assumptions in optimization research are completely wrong.
(paper out soon)

7601

Reposted by mr. TIM

Sonnet 5.5 is out! But if you’re on Amazon, only if you’re prepared to let them run it anywhere in the world. Not awesome for the many companies that enforce data sovereignty constraints. docs.aws.amazon.com/bedrock/late...

docs.aws.amazon.com

Anthropic - Amazon Bedrock

The following Anthropic models are available in Amazon Bedrock:

141

softmacs @softmacs.bsky.social · 28/09/2026

progressive disclosure is when you put your pronouns in your bio

1350

Reposted by mr. TIM

Anthropic is just serving up a single model with a pay-what-you-want pricing scheme really

0173

Claude 5.5 Sonnet is live and it’s roughly Opus 5.5 but cheaper www.anthropic.com/claude-sonne...

A benchmark comparison table titled "Claude Sonnet 5.5" comparing four models: Sonnet 5.5, Sonnet 5, Opus 5.5, and GPT-6 Sol.
* Agentic coding (Terminal-Bench 4.0): Sonnet 5.5: 70.6%; Sonnet 5: 10.3%; Opus 5.5: 66.4%; GPT-6 Sol: N/A.
* Agentic coding (FrontierCode 1.1 Main): Sonnet 5.5: 46.2% (Max) / 52.1% (Xhigh); Sonnet 5: 42.4%; Opus 5.5: 54.4%; GPT-6 Sol: 49.3%.
* Agentic coding (CursorBench 4.0): Sonnet 5.5: 55.5%; Sonnet 5: 34.1%; Opus 5.5: 57.8%; GPT-6 Sol: N/A.
* Knowledge work (GDPval-AA v2.1): Sonnet 5.5: 1844; Sonnet 5: 1449; Opus 5.5: 1846; GPT-6 Sol: 1487.
* Knowledge work (AA-Briefcase v1.1): Sonnet 5.5: 1811; Sonnet 5: 1359; Opus 5.5: 1822; GPT-6 Sol: 1483.
* Multidisciplinary reasoning (Humanity's Last Exam with tools): Sonnet 5.5: 64.5%; Sonnet 5: 54.9%; Opus 5.5: 67.7%; GPT-6 Sol: N/A.
* Computer use (OSWorld 2.1 partial): Sonnet 5.5: 80.1%; Sonnet 5: 57.0%; Opus 5.5: 81.8%; GPT-6 Sol: N/A.
* Visual chart recognition (Chartography no tools): Sonnet 5.5: 61.6%; Sonnet 5: 15.6%; Opus 5.5: 64.4%; GPT-6 Sol: 53.6%.
Footnotes provide methodology notes regarding evaluation settings, Artificial Analysis pre-release testing details, and recent bug fixes affecting GPT-6 Sol benchmark scores.

A line graph titled "Agentic coding by effort level" on the CursorBench 4.0 benchmark, plotting Score (%) on the linear y-axis (20% to 60%) against Cost per task in USD on a logarithmic x-axis ($0.5 to $10+).
The chart compares four models:
* Sonnet 5.5 (blue line with labeled effort levels): Starts at "Low" (~$0.50, 36%), moving through "Med" ($0.70, 39%), "High" ($1.70, 48%), "Xhigh" ($3.80, 53%), and "Max" ($9.50, ~55.5%).
* Opus 5.5 (orange line): Tracks closely above Sonnet 5.5 at higher cost points, spanning from ~$1.20 per task (~44%) up to ~$12.50 per task (~58%).
* Sonnet 5 (grey line): Shows lower accuracy relative to cost, ranging from ~$0.90 per task (~25%) to ~$8.00 per task (~42%).
* GPT-5.6 Sol (light green line): Represents the lowest trajectory, ranging from ~$1.40 per task (~24%) to ~$7.00 per task (~34%).
Footnote: "CursorBench 4.0 does not report GPT-6 Sol performance publicly, so we report GPT-5.6 Sol here."

2020623

this statement: “{OpenAI|Anthropic} released cheaper model X because they are short on compute” does not make sense if you believe in the Jevons Paradox

6511

OpenAI’s dev day is on Tuesday. They have “too much to launch”, meaning they’ll likely spill over the launches into Monday & Wednesday. I think we can expect another GPT-6 model, possibly Aeon (highly persistent), along with new products that make use of the persistent behavior

3260

given the existence of things like embeddings and PCA, we should be able to plot the vibe shift here “AI go rogue” is a very different conversation from “stochastic parrot”. things are shifting

1170

seems inevitable tbqh software engineer is (actually) hard, and too many people never operated at the level you have to operate now so yeah, laziness..

4551

seems likely that Sol-6 is a much smaller model than Sol-5.6 also seems likely that Opus-5.5 is smaller (a lot?) than Opus-5 i think labs are finding ways to replicate a lot of that big model feel in smaller models, which is a good sign for that cognitive core dream

6821

i think GPT-6 Sol is a good model, but i don’t like coding with it, at least not directly

5110

platonic forms enter the chat

A cartoon depicting two stick figures standing behind a chain-link fence at an airport as an airplane takes off in the background. The stick figure on the left points at the plane and says in a speech bubble, "It's not really flying. It doesn't have feathers." The stick figure on the right looks on thoughtfully while imagining a parrot perched on a branch inside a thought bubble.

4582

one of the dangers in RSI is you don’t have to make models that people like when you sell models, they have to be aligned enough for people to actually buy them with RSI, the goal is entirely different

2233

Eugene Vinitsky 🍒 @eugenevinitsky.bsky.social · 26/09/2026

i would say, in most contexts it is probably unwise to focus on risks you’ve never experienced e.g. in software engineering i’ve developed a bit of a gag reflex whenever an engineer is solving a problem that doesn’t exist yet trouble is the severity of x-risks puts them in a unique position

Andy Masley • @AndyMasley
X.com
Are there any other contexts where people are asked to never consider or work on risks that could pop up in 5 years because they "distract from the real harms in the present"?

6571

Reposted by mr. TIM

"These people are really annoying and/or weird" is not actually an argument against their point

9945

this by no means solves alignment, but a whole lot of misaligned behaviors stem from reward hacking, so this is a very big deal

5464

RLMs & Program Agents I'm experimenting with this idea, Program Agents, inside of DeepSeek Harness (DSH) They're like RLMs, except without the LLM. Agents are writing such large blocks of code, what if they just never exited? github.com/tkellogg/dsh...

Diagram illustrating the architecture of an RLM (Representation Learning Model / Recursive Language Model architecture) system:

* At the top, a container labeled **RLM** holds a central block labeled **Parent RLM**.

* Below it is a container labeled **Program Agents**, which houses a central block labeled **Program Agent**.

* Two curved directional arrows connect **Parent RLM** and **Program Agent**:
* An arrow pointing down from **Parent RLM** to **Program Agent** labeled **edit, restart**.

* An arrow pointing up from **Program Agent** to **Parent RLM** labeled **report error**.

* At the bottom row, four separate blocks labeled **Subagent** are connected via downward-pointing arrows originating from the **Program Agent** block.

4292

New from ArtificialAnalysis: The price of compute per KG set to surpass the price of cocaine in 2029

Line graph titled "The Compute-Cocaine Crossover: $/kg, 2024-2030E" comparing the price per kilogram in USD (logarithmic scale) of NVIDIA compute hardware against illicit drugs and gold over time.
Key elements in the chart include:
* Subtitle: "Chart 1. Compute is the only commodity whose $/kg rises every generation while everyone else's supply chain gets more efficient."
* Gold: A constant horizontal yellow line at $138k/kg.
* Cocaine (US wholesale): A white line sloping downward from around $35k/kg in 2024 to about $15k/kg by 2030.
* NVIDIA NVL72 Racks (Actual & Projected): A green line starting at $3.2k/kg in 2024 for the GB300 NVL72 ($5M / 1,580 kg), rising to $5.5k/kg in 2026 for the Vera Rubin NVL72 ($8.8M), and continuing upward as a dashed projection until it crosses above the cocaine curve around 2029–2030 ("Crossover SemiAnalysis est.").
* Fentanyl (Wholesale): A constant flat red line at $3.5k/kg.
* Cannabis Flower (US wholesale spot): A constant flat dark-green line at $2.4k/kg.

5404

is openai releases Jevstra at dev day on tuesday, that’ll be quite a statement about the level of RSI we’re at

1110

looks like codex usage just reset

6220