CEO - founder of complicated things inc - board member - Singularity U Grad - Writer + Speaker
Joined June 2008
- Tweets13K
- Following744
- Followers425
- Likes638
Pinned Tweet
Other writing as well: complicatedthings.com/words-…
ɔıɔıɾnɹƃ ɐɥsɐs retweeted
$6.75 in frontier-model cost.
$0.46 with GoodEnough Auto.
Same illustrative workload: 1M input tokens and 250K output tokens.
That is a stated 93% cost difference.
But the cheap number is not the point.
The point is whether the output clears your real task.
A lower token price does not prove a model is right for your work. GoodEnough's own advice is to test on your prompts, inspect the result, and keep the model that meets the need.
So don't ask, "What is the smartest model?"
Ask:
What did this task cost?
Did the answer pass review?
Would I pay the premium again?
Run 100 real prompts. Compare the output and the invoice.
Then let the numbers decide whether your default model has earned its price.
ɔıɔıɾnɹƃ ɐɥsɐs retweeted
Stop routing every prompt to the flagship.
Let Auto choose:
• GPT-OSS for general chat
• Qwen3 Coder Next for code
• Qwen3 VL for images
Then inspect the `model` field in every response.
That field is your workload audit.
Run 100 real prompts. Count the model mix. Rerun evals by task. Compare output quality and cost.
If most requests do not need the flagship, paying for it by default is not a strategy. It is a tax.
GoodEnough Auto returns the selected model, so you can measure the routing instead of trusting the label.
ɔıɔıɾnɹƃ ɐɥsɐs retweeted
Coming soon to Good Enough AI! 85% off your AI search bill.
api-docs.deepseek.com/guides…
ɔıɔıɾnɹƃ ɐɥsɐs retweeted
One model for every request is an expensive habit.
Let Auto route the work:
general chat -> GPT-OSS
code -> Qwen3 Coder Next
images -> Qwen3 VL
GoodEnough's router uses a learned classifier plus deterministic filters. The response includes the `model` field, so you can inspect what actually handled each request.
Run 100 real prompts:
1. Log the model field.
2. Group requests by task.
3. Compare quality and spend.
Auto is listed at $0.23 per 1M input tokens and $0.92 per 1M output tokens.
Stop debating which model wins.
Measure which model your requests actually need.
ɔıɔıɾnɹƃ ɐɥsɐs retweeted
Everyone is looking at model routing like it is five model choices.
The more interesting part is that it can become one shared decision system.
With an OpenAI-compatible API, your application can send requests to one endpoint. Auto then selects a model for each request and returns which model it chose.
A coding task can go to Qwen3 Coder Next. An image request can go to Qwen3 VL. General chat can go to GPT-OSS.
That turns one API key into a model fleet.
The upside is obvious:
- fewer provider integrations
- fewer model-selection branches in application code
- the option to use function calling and tools through the same interface
- a cost decision made per request instead of per product
The blast radius is just as real:
- your bearer token can reach every model your account exposes
- a routing rule becomes a production dependency, even when it looks like a pricing setting
- tool calls need the same approval boundaries whether the request landed on a cheap model or an expensive one
- rate limits and account credit are separate constraints, so "the API is up" does not mean "the workload can continue"
- if you do not record the selected model and token cost, you cannot explain the invoice or the output mix
The breakthrough is not that you can call several models.
It is that they can look like one provider inside your real application.
That changes the job from choosing a flagship once to designing routing, permissions, fallbacks and measurement.
Do not compare models on a leaderboard and stop there.
Run 100 real prompts through the route you plan to ship. Check which model handled them, what the outputs cost, and where a human approval still belongs.
ɔıɔıɾnɹƃ ɐɥsɐs retweeted
A $2,000 AI bill can become an estimated $134.62.
GoodEnough's calculator gets there using a 75% input / 25% output mix.
That is a 93.3% difference.
But don't trust the calculator. Trust your workload.
Take 100 production-like prompts:
- support replies
- code changes
- extraction
- long-document questions
- image requests
Run them through Auto. It routes each request to a model, and the response shows which model it selected.
Then compare:
1. Did the answer pass your eval?
2. How many tokens did you use?
3. What would the bill have been at your current rates?
GoodEnough Auto is listed at $0.23 per 1M input tokens and $0.92 per 1M output tokens.
The point isn't that cheaper is always better. GoodEnough explicitly says to test your own work and keep the model that earns its place.
The point is that "we need the flagship" is a claim your 100 prompts can disprove.
Your bill is probably hiding a routing problem.
ɔıɔıɾnɹƃ ɐɥsɐs retweeted
A $2,000 AI bill can be about $134.62.
Here's the calculation:
75% input / 25% output
Frontier reference:
• $3 per 1M input tokens
• $15 per 1M output tokens
GoodEnough Auto:
• $0.23 per 1M input tokens
• $0.92 per 1M output tokens
Weighted frontier rate:
0.75($3) + 0.25($15) = $6
Weighted Auto rate:
0.75($0.23) + 0.25($0.92) ≈ $0.40
$2,000 × ($0.40 ÷ $6) ≈ $134
The calculator's estimate is $134.62, or 93.3% less. The small difference comes from rounded public rates.
Before accepting your next frontier-model bill, run this calculation on your own input/output mix.
Your workload may be expensive. Your default probably is.
ɔıɔıɾnɹƃ ɐɥsɐs retweeted
🚨Final nail in the coffin of the AI IPOs? 🚨
“The ten highest-quality models carry 14% of developer tokens, down from 55% last year, as cheaper, "good enough" models win on volume.”
theinformation.com/articles/…
ɔıɔıɾnɹƃ ɐɥsɐs retweeted
Open source is the only way to protect your projects.
GLM-5.3 has helped defend 389 open-source projects, with 4,249 potential vulnerabilities found so far.
OpenVuln is still running. The service remains free, and findings go privately to maintainers.
huggingface.co/spaces/zai-or…
ɔıɔıɾnɹƃ ɐɥsɐs retweeted
Regulators Find Their Situational Awareness: Fed, BoE Probe Bank Exposure To Jane Street After AI Fund Meltdown zerohedge.com/markets/regula…
ɔıɔıɾnɹƃ ɐɥsɐs retweeted
The group behind such hits as: “We’re All Going To Die” and “Regulate Me Now” are building a wet lab in SF.
I do not recommend this.
ANTHROPIC BUILDS BIOLOGY LAB TO PUSH AI INTO DRUG DEVELOPMENT
Anthropic has quietly established a Bay Area wet lab, moving its life-sciences ambitions beyond computer simulations and into physical biological experiments.
The company wants Claude to direct laboratory robots with limited human intervention, potentially accelerating research into rare and traditionally “undruggable” diseases.
Anthropic says it will focus on preclinical research rather than clinical trials, avoiding direct competition with pharmaceutical companies.
ɔıɔıɾnɹƃ ɐɥsɐs retweeted
We'll show you the cheapest model that passes your evals.
Find your model & sign up:
itsgoodenough.ai/?utm_source…
ɔıɔıɾnɹƃ ɐɥsɐs retweeted
The cheap model is also the fast one this month.
GPT-6 Astra streams 59 tokens a second on OpenAI's API (Artificial Analysis, Sept 2026). Groq serves Llama 3.3 70B at 312.
5x the output speed, and the price gap runs the other way: $10/$50 per million vs $0.10/$0.32 on public host lists. Under 1% of the cost, blended.
If a batch job is crawling and the bill is climbing, you picked the model that loses on both axes.
ɔıɔıɾnɹƃ ɐɥsɐs retweeted
🦔Anthropic published an economic model this week on how AI could affect the US economy by 2030. They ran three scenarios. In the modest one, AI has about the same impact as the internet. In the substantial one, GDP growth doubles but knowledge worker wages go flat. In the extreme one, GDP grows 15% a year, unemployment spikes beyond recessionary levels, knowledge worker wages fall over 10%, and the share of GDP that goes to workers drops from 60% to 45%.
The company's conclusion is that "the challenge is making sure the benefits are broadly shared."
My Take
The economy already swaps $120,000 knowledge work careers for $35,000 service jobs. The Dallas Fed showed 13% fewer 22-to-25 year olds employed in AI-exposed fields than 2022. Corporate profits are at highs we haven't seen since World War II and workers are taking home less of the pie than any time since the 1950s. Anthropic's own model says AI makes that worse.
They're also about to IPO and they want pension funds to be able to buy their debt. So a paper that says "we're thinking about this responsibly" is a pretty useful thing to have right now. The paper says someone should make sure the gains get shared. I've been around long enough to know that when a company says "someone should fix this," they don't mean themselves. Manufacturing towns heard the same promise 20 years ago and they're still waiting.
Hedgie🤗
anthropic.com/institute/econ…
ɔıɔıɾnɹƃ ɐɥsɐs retweeted
We're hiring a co-founder: CEO, Commercial.
One job: revenue.
Run the pipeline yourself and close the first customers. Write the sales playbook by doing the selling. Hire the first reps once it repeats. Own pricing a CFO can budget around. Report the numbers to the board, then lead the raise when the proof is in.
You've done this before: taken a B2B company from early traction to $1M to $10M ARR. Founder-led sales. Mid-market buyers.
Co-founder equity + salary.
This! The world is waking up to data access as being the real value multiplier on model improvements. It's so obvious ... it's the live context that matters
Pretraining progress seems to be coming mostly from data improvements.
@who_is_jerbear and I pretrained combinations of year-representative open model recipes and data corpuses across 2019 to 2025 at various small scales.
Data improvements contributed 3.24x as many compute multipliers as model improvements did (12.0x vs 3.7x).
And the gains stack independently - a better dataset helps every architecture about equally, and vice versa.
Here are full results, plus what we think this means for the future of AI progress:
dwarkesh.com/p/pretraining-p…
ɔıɔıɾnɹƃ ɐɥsɐs retweeted
Your AI bill is 70-90% habit.
Cut the cost & join the waitlist:
itsgoodenough.ai/?utm_source…
ɔıɔıɾnɹƃ ɐɥsɐs retweeted
The biggest AI discount sits on your provider's own price page. GPT-5.4: $5.625 blended. GPT-5.4 nano: $0.4625. 12x, same API key.
Opus 4.8 to Haiku 4.5 is 5x. Gemini 3.1 Pro to Flash-Lite is 8x.
Nobody sells you the cheap tier. You have to eval your way into it.
ɔıɔıɾnɹƃ ɐɥsɐs retweeted
We'll show you the cheapest model that passes your evals.
Find your model & join the waitlist:
itsgoodenough.ai/?utm_source…
ɔıɔıɾnɹƃ ɐɥsɐs retweeted
56% off AI coding costs for a 2% quality drop.
AT&T just took that trade in public: 40% of internal AI usage on open models, headed for 60 to 70%, frontier spend held flat, per The Information.
Nobody gets fired for defaulting to the flagship. AT&T just published what the default was costing.